How do available AI agent memory layers compare for developer workflows?
State of AI Agent Memory 2026. Facts read 5 to 13 October 2026 and re-read on 9 November 2026.
We build Mnemoverse, one of the systems in this report. Each benchmark number carries its recipe and source, each absence claim its pages and date, and the section where Mnemoverse comes out worst is last before the sources.
TL;DR
- In October 2026, 23 products qualify as AI agent memory layers under the rule below; 17 are compared on the axes.
- Outcome feedback. A documented call marking a recalled memory helpful or wrong exists at Mem0 (hosted only), MinnsDB and Mnemoverse; Hindsight marks facts invalid; Cognee and Supermemory route ratings or reviews into retrieval; MemNexus records selections; Memori takes feedback without a memory id.
- Across clients. Mem0, Zep, Letta, Cognee, Supermemory and Mnemoverse all document one memory shared by several tools.
- No vendor account. Mem0, Letta, Cognee and Supermemory document a path, Zep through Graphiti; Mnemoverse does not.
- Benchmarks. Vendor scores differ in questions, memory depth, judge and answering model, so they do not rank systems.
- Standards. No ratified agent memory standard was found in October 2026.
An AI agent memory layer is a product or library whose main job is an agent's memory, which a developer connects through an API, an SDK, an MCP server or a plugin.
An outcome input is a documented API, SDK method or MCP tool with which an agent or a person, after use, marks a specific recalled memory as helpful or wrong. A label on an agent step, a rating of a session answer, a review of an inferred memory, a rating of a rule, and a call that records a rating without applying it are not counted here; they appear in a separate column as the nearest mechanism.
An account-free path is a documented way to run the memory layer on your own machine or servers with no account at the vendor.
Which agent memory systems this report covers
A system is described here if one of two conditions holds:
- it is one of the six systems of the comparison Best Persistent Memory APIs for AI Agents: Mnemoverse, Mem0, Letta, Zep (with Graphiti), Cognee and Supermemory; or
- at least one of 28 third-party comparison pages, the pages AI answer engines cited in answers to agent-memory buyer questions in a citation export of 28 September 2026, names it as a system in a list or table, and it is a memory layer by the definition above. A vendor's own comparison page counts: that is how EverOS (on the EverMind blog) and MemNexus (on its own blog) qualify. The 28 pages are listed under Sources.
Agent frameworks with built-in memory (LangGraph, LlamaIndex, CrewAI, AutoGen, Mastra, Cloudflare Agents), databases and vector stores (Pinecone, Qdrant, Weaviate, Chroma, Milvus, Neo4j, Redis as a database), search and RAG engines (LightRAG, Glean, Shaped) and the memory built into assistant platforms do not count. Platform memory has its own section.
On 13 October 2026 the rule gives 23 systems: 6 by the first condition and 17 by the second. The six and 11 of the second group, 17 in all, are described on the axes below; the 11 are LangMem (named on 17 of the 28 pages), Hindsight (7), MemoryLake (2), and MemMachine, EverOS, TencentDB Agent Memory, Memori, Redis Agent Memory, Sentra, MinnsDB and MemNexus, each named as a product on one page. Six are named by the rule and classified from their own sites on 13 October 2026; their documentation was not read for this report, so they are not described on the axes:
- Membase: "Membase automatically brings your personal context from your daily agents and apps." (membase.so)
- Memory Store (MemStore): "Shared memory for your team's agents" (memory.store, read 9 November 2026)
- mem9: it "gives the agents you use one shared memory layer" (mem9.ai)
- Kumiho: "The memory layer AI agents can trust." (kumiho.io)
- Memora: "Give your AI agent a memory that lasts." (memora.sh)
- Weaviate Engram: "Engram is a memory server for LLM agents and applications.", a service separate from the Weaviate database (docs.weaviate.io/engram)
Considered and not included, each for its own vendor's description of its main job (read 11 and 13 October 2026; PuppyOne and Coworker re-read 9 November 2026):
- Microsoft Kernel Memory: its README calls it "an archived research project" whose service is "specialized in the efficient indexing of datasets", with support for retrieval augmented generation, which places it with search and RAG engines.
- Graphlit: "Graphlit is the context layer underneath agents.", and its sunset notice says "Graphlit is not accepting new customers or self-service signups."
- PuppyOne: "puppyone is the folder your agents live in." (puppyone.ai)
- Perplexity Brain: a memory feature of Perplexity Computer; "Brain is rolling out today to Max and Enterprise Max subscribers in Research Preview." That is platform memory.
- Coworker: "Coworker is the sovereign context layer for enterprise AI." (coworker.ai), with organizational memory as one part of the platform.
Sentra and Coworker sit near the same border, and one rule decides both: the main job in the vendor's own words. Sentra calls itself "the managed organization memory, the company brain a company's teams and AI agents share", so it is in; Coworker calls itself an agent platform, so it is out. Letta also publishes the experimental AI Memory SDK (letta-ai/ai-memory-sdk, Apache-2.0), a wrapper around Letta agents, last pushed on 4 November 2025; it is counted with Letta.
Does a recalled memory's outcome change what comes back next?
This is the axis where the systems differ most, and where the wording on vendor pages is easiest to over-read. The table applies the definition of an outcome input given above. Each row is bounded by the pages read on the date in its last column.
| System | Outcome input, by the definition | Nearest mechanism, where there is no input | What the vendor says about the effect on retrieval | Pages read |
|---|---|---|---|---|
| Mem0 | yes, on the hosted platform only: a feedback call (POSITIVE, NEGATIVE, VERY_NEGATIVE); listed as not available in Mem0 Open Source; the documented hosted MCP tool list has no feedback tool (the live list was not read) | "used to improve the accuracy of the memories and search results"; no ranking formula, weight or parameter is given | 249 docs.mem0.ai pages, 8 Oct 2026 | |
| MinnsDB | yes, in two steps: REST events MemoryUsed {query_id, memory_id} and Outcome {query_id, success} | docs: "Final outcome of the learning cycle"; API reference: "close the feedback loop"; the effect is described only in code | docs.minns.ai and API reference, 13 Oct 2026 | |
| Mnemoverse | yes: REST "Report outcome (success/failure) for memories", MCP memory_feedback, SDK feedback | "Ratings re-rank what comes back next." | Mnemoverse docs, 7 Oct 2026 | |
| Hindsight | no input for "helped"; a stored fact can be marked invalid by id with an optional reason (REST PATCH, MCP tool invalidate_memory) | curation of a stored fact (edit, invalidate, restore); ranking boosts for recency, temporal proximity and proof count | an invalidated fact disappears from recall, reversibly | llms-full and source, 8 Oct 2026 |
| Cognee | no, by the definition: a 1 to 5 rating attaches to a session answer (SDK cognee.session.add_feedback, CLI, REST), not through an MCP tool | a rating of a session answer, which improve() carries to the graph elements the answer was built from | "In practice, this means highly rated answers can make their source graph elements more influential later, while poorly rated answers can reduce their influence."; "To make feedback influence future retrieval, run improve() with the relevant session_ids."; the influence defaults to 0.0 (DEFAULT_FEEDBACK_INFLUENCE) | 567 docs.cognee.ai pages, 8 Oct 2026 |
| Supermemory | no | review of memories the engine inferred, which are "flagged as inferred (isInference: true) until someone confirms them"; an approved one "Stays in search, no longer flagged as inferred", a declined one is "Removed from search entirely" | the path from review into search is documented | docs, OpenAPI v3 to v5 and MCP source, 8 Oct 2026; review page re-read 9 Nov 2026 |
| MemNexus | no, by the definition | REST POST /api/memories/search/feedback and the SDK record whether a result was selected or not selected, per record, without an MCP tool; a CONTRADICTS relationship to a record drops it from default search | the endpoint is described as "Record behavioral feedback on a search result" | 52 docs pages and SDK 1.77.11, 13 Oct 2026; endpoint re-read 9 Nov 2026 |
| Memori | no: memori_feedback takes free text and an optional positive or negative sign, with no memory id | free-text feedback on memory behavior | the MCP skill says "Feedback improves memory extraction quality, recall relevance, and summary accuracy."; the integrations describe it as feedback to the Memori team | 116 site and docs pages and the repositories, 13 Oct 2026 |
| Letta | no | a positive or negative label on an agent step (PATCH /v1/steps/{step_id}/feedback), used as a filter when listing steps; feedback given in plain language makes the agent rewrite its own memory | no path from the step label into retrieval is documented | 635 docs.letta.com pages, 8 Oct 2026 |
| Zep | no | Fact Ratings rated facts at extraction time, and "Fact ratings are being deprecated entirely" | 361 help.getzep.com pages, 8 Oct 2026 | |
| EverOS | no, in the docs and the open-source code | a pre-release cloud SDK (1.2.0-rc1) adds feedback(memory_id, rating), documented as "Recording is not applying" | 104 docs.evermind.ai pages, 8 Oct 2026 | |
| TencentDB Agent Memory | none on the pages read | the agent repairs a skill with the skill_patch tool; session review revises skills a session disproved; the docs give a correction template for chat; the vendor's article names down-weighting (降权) as a principle, with no API for it found | repository branch feat/server_team and cloud docs, 9 and 10 Oct 2026 | |
| MemoryLake | none on the pages read | curation (edit, forget with history kept, conflict resolution); an up or down rating on an A2A task, not on a memory | "Edits take effect immediately in future retrieval." | 270 English docs pages and llms-full, 13 Oct 2026 |
| Redis Agent Memory | none, in Redis Iris (14 OpenAPI operations) or in the open V0 server | edit and delete by id; in V0, a recency boost | redis.io docs, OpenAPI and V0 source, 13 Oct 2026 | |
| Sentra | none on the pages read | a person corrects facts in the Memory browser | 4 docs pages, llms.txt, OpenAPI 2026.10.13, 13 Oct 2026 | |
| MemMachine | none on the pages read | 138 docs sections and main branch, 8 Oct 2026 | ||
| LangMem | none on the pages read | 22 docs pages and repository, 8 Oct 2026 |
Read across the table in one line: by the definition, an outcome input is documented by Mem0 (hosted platform only), MinnsDB (two REST events) and Mnemoverse; Hindsight documents marking a single stored fact invalid; Cognee and Supermemory document a path into retrieval from answer ratings and from review of inferred memories; MemNexus records whether a search result was selected, for measurement; Memori takes free-text feedback with no memory id; Letta labels agent steps; and Zep, EverOS (outside a pre-release SDK), TencentDB Agent Memory, MemoryLake, Redis Agent Memory, Sentra, MemMachine and LangMem document none on the pages read.
The word for the negative direction differs by vendor, and no cross-vendor count of it is attempted here. Supermemory's review page describes the negative direction as removal: a declined inferred memory is "Removed from search entirely" (read 9 November 2026). Cognee describes the negative direction in other words: "poorly rated answers can reduce their influence". TencentDB's article names down-weighting (降权) as a principle. The Mnemoverse MCP tool text (package 0.15.0, the same in 0.17.1 on 9 November 2026) says "negative feedback lowers it so other memories out-rank it".
Does one memory follow you across clients?
All six systems of the comparison page document a route by which one memory is reached from more than one client. What differs is the route.
- Mem0: "The shared Mem0 editor plugin now covers Kimi Code alongside Claude Code, Cursor, Codex, and Antigravity, running on the same scripts, skills, and memory bank, so context written in one editor is available in the others." (changelog entry of 13 August 2026, read 5 October 2026)
- Zep: the Context MCP Server shares one user graph between an in-house agent and personal MCP clients such as Claude, ChatGPT, Claude Code, Codex and Cursor; it is on every plan through per-account MCP seats (read 8 October 2026).
- Letta: the hosted MCP server, for clients that send custom bearer headers ("OAuth-only clients are not supported yet."), the Agent Client Protocol, and the same agent used from several computers (read 8 October 2026).
- Cognee: its MCP server keeps writes apart per client dataset by default, but recall without named datasets searches every dataset the user can access (read 8 October 2026).
- Supermemory: its plugins for Claude Code, Codex and OpenCode derive the same container tag for a repository, so memories written from one are shared with the others (README, read 8 October 2026).
- Mnemoverse: any key or sign-in from the same account reaches the same memory, and rooms (beta) share memory across accounts.
Beyond the six, on the pages read 13 October 2026 unless noted:
- MemoryLake: "Your personal workspace becomes the single memory all your AI tools share."; a remote MCP server with OAuth2 whose tools only read ("These tools read and analyze; they do not modify or delete your documents and memories."), plus plugins and a client harness.
- Memori: "Memories are shared across providers since they're linked to entities, not providers."; a hosted MCP server with attribution headers, documented for Cursor, Claude Code, Codex, Warp and Antigravity.
- MemNexus: one remote MCP server; the CLI's
mx setupconfigures Claude Code, Claude Desktop, GitHub Copilot in VS Code, Cursor, Codex CLI, OpenCode and ChatGPT Desktop. - Sentra: one organization graph through a remote MCP server, with one-click connection for ChatGPT, Claude, Claude Code and Codex, VS Code through mcp-remote, Cursor with a key, and Windsurf and Zed through mcp-remote.
- Redis Agent Memory: a store reached over REST and SDKs. On MCP the vendor's pages disagree: the self-managed page lists "The MCP streamable-HTTP endpoint for a store.", and the Google ADK page says "The redis-agent-memory backend does not expose an MCP endpoint, on Redis Cloud or self-managed; use the REST tools above."
- MinnsDB: one server or cloud project reached over REST and SDKs; no MCP server is described on docs.minns.ai or in the MinnsDB repository.
- TencentDB Agent Memory: a proxy for seven kinds of client and plugins for OpenClaw and Hermes; on the default branch a memory is private, team or restricted (access lists by user, role and agent), and "New Chat Memory and Skills are private by default. Sharing is an explicit action, not a default leak." (README of branch
feat/server_team, read 9 October 2026). - Hindsight: a portable plugin connects to the hosted endpoint over streamable HTTP with a bearer header, and its Coding Agents integration listed 22 harnesses on its page on 8 October 2026.
- MemMachine: an OpenClaw plugin and an agent skill installable for Codex, Claude, Gemini and OpenCode; its MCP server exposes add, search and delete (read 8 October 2026).
- EverOS: EverMe plugins; their MCP configuration was not read.
- LangMem: a library over a LangGraph store; no MCP server or host plugin is documented on the 22 pages read on 8 October 2026.
Can you run it without a vendor account?
| System | Account-free path | How, in the vendor's terms |
|---|---|---|
| Mem0 | yes | Mem0 Open Source; on-prem deployment of the hosted platform is a Custom-plan feature (pricing page, read 9 Nov 2026) |
| Letta | yes | Letta Code runs locally without a Letta account (Apache-2.0); since 0.32.16 it starts against Letta Cloud by default (read 8 Oct 2026) |
| Zep | only through Graphiti | Zep's FAQ says Zep Community Edition "is deprecated and no longer supported"; the account-free path is Graphiti (Apache-2.0), its MCP server and a graph service published as the Docker image zepai/graphiti (read 8 Oct 2026) |
| Cognee | yes | the open-source engine, run locally or on your own stack (read 13 Oct 2026) |
| Supermemory | yes, as a closed binary | the self-hosted server is a closed binary ("The server binary is not open source") that runs without a Supermemory account and needs one LLM provider key; telemetry is on unless SUPERMEMORY_DISABLE_TELEMETRY=1 is set. Separately, the pricing page says "Self-hosted deployments are available on Scale and Enterprise." (read 8 and 13 Oct 2026) |
| Mnemoverse | no | Enterprise customers can self-host by agreement; no self-hostable build is published |
| Hindsight | yes | "Hindsight is MIT licensed with no usage limits, no telemetry, and no restrictions."; the Enterprise card on the pricing page, reworded on 13 October 2026, reads "Hindsight in a private cloud, your own AWS, Google Cloud or Azure account, or on premises, with enterprise security and contracted support." |
| TencentDB Agent Memory | yes | MIT; storage and retrieval run locally; extraction calls an LLM at an address the user sets; in v0.3.6 that address defaulted to Tencent Cloud LKE (read 9 and 10 Oct 2026) |
| MemMachine | yes | Apache-2.0; the pricing page's open-source plan reads "Self-hosted on your own infrastructure." (read 13 Oct 2026) |
| EverOS | yes | Apache-2.0, a "local-first memory runtime"; "One OpenRouter API key is enough to start EverOS" (read 8 Oct 2026) |
| LangMem | yes, as a library | MIT, over the LangGraph store; no page about running a LangMem server was found on the pages read, and the managed service announced in its launch post of 18 February 2025 was not checked (read 8 Oct 2026) |
| MinnsDB | yes | AGPL-3.0; "MinnsDB needs a Qdrant vector store." (README, read 13 Oct 2026) |
| Redis Agent Memory | no, for Redis Iris; yes, for the open V0 server | self-managed Iris needs a license: "A license key is required to deploy."; the open predecessor runs without an account under Apache-2.0, and Redis keeps it "as an open research artifact, not as the supported production distribution." (read 9 Nov 2026) |
| Memori | storage yes; extraction calls the vendor's API, without an account but rate-limited | the Apache-2.0 SDK stores memory in your own database without an account; extraction (Advanced Augmentation) calls the vendor's hosted API, without an account but rate-limited (README and FAQ, read 13 Oct 2026) |
| MemoryLake | no | a managed service; on-premise deployment is announced for enterprises (changelog 1.6.0); no self-hostable build is published (read 13 Oct 2026) |
| MemNexus | no | "Self-hosting option" is listed for Enterprise; no self-hostable build or source is published (read 13 Oct 2026) |
| Sentra | none on the pages read | deployment as "Cloud, isolated VPC, air-gapped on-prem" by contract, per the vendor's comparison table (read 13 Oct and 9 Nov 2026) |
Plugins, hosts and the Agent Plugins format
The portable Agent Plugins v1 format is carried by Mnemoverse (claude-plugin), Mem0 (mem0-agent-plugin), Zep (an mcp.json with the agent-plugins.org schema) and Hindsight (hindsight-integrations/agent-plugin), read 8 October 2026. On the host side, Cursor's documentation says "A plugin that follows the Agent Plugins specification loads in Cursor.", and adds that Cursor does not expand the standard's ${PLUGIN_ROOT} and ${PLUGIN_DATA} variables in mcp.json (read 8 October 2026).
How a plugin reaches memory differs, and no system is ranked on it:
- Mnemoverse:
mcp.jsonwith streamable HTTP tohttps://mcp.mnemoverse.com/mcp. - Zep: HTTP to
https://api.getzep.com/mcpin both.mcp.jsonandmcp.json. - Hindsight: streamable HTTP with a bearer header to
https://api.hindsight.vectorize.io/mcpin the portable plugin and the Grok Bot plugin; its Claude Code plugin runs a local MCP server. - Supermemory: the Claude Code plugin runs a local proxy (
hooks/mcp-proxy.js) tohttps://mcp.supermemory.ai/mcp; the Cursor plugin runs a local CLI. - Mem0: the plugin manifests declare only local stdio MCP servers; the hosted server at
mcp.mem0.aiis a separate client configuration. - Cognee: no MCP server in the plugin manifests; Cursor and Gemini CLI are documented through MCP configuration.
- Letta: one host plugin, Claude Subconscious for Claude Code, which its README calls a demo app "not intended to be used in production".
- TencentDB Agent Memory: the OpenClaw and Hermes plugins talk to its Memory Gateway over HTTP.
- MemMachine and EverOS: the MCP configuration of their plugins was not read.
Hosts by vendor, read 8 October 2026: Mem0's repository holds plugin directories for ten agent hosts (Antigravity, Claude Code, Codex, Cursor, DeepSeek Harness, Hermes, Kimi Code, OpenClaw, OpenCode, Pi Agent) plus a portable package. Supermemory maintains seven host plugin repositories (Claude Code, Cursor, Codex, OpenCode, OpenClaw, Hermes, Muse). Zep carries manifests for Claude, Codex and ChatGPT Work, Cursor and the Agent Plugins format, and sends VS Code, GitHub Copilot and Gemini CLI to its remote Context MCP Server. Cognee ships manifests for Claude Code, Codex and Antigravity, and its VS Code extension cognee.cognee-vscode is on the Visual Studio Marketplace (version 0.1.0, last updated 23 July 2026). Letta's Cursor and VS Code route is its hosted MCP server. Hindsight's Coding Agents integration targets Antigravity CLI (agy), which Hindsight describes as the renamed Gemini CLI. MemMachine ships an OpenClaw plugin and an agent skill. Mnemoverse's hosts with the status Shipped or Verified live are listed in its section.
What agent memory layers cost
Prices as each vendor's page stated them when re-read on 9 November 2026, the date of every price in this table. Plans change; check the vendor's page before deciding. Open-source and self-hosted editions are in the self-hosting table above.
| System | Free or open | Paid plans | Notes |
|---|---|---|---|
| Mem0 | Free: unlimited end users; for the first 30 days 10,000 adds and 1,000 retrievals a month with model calls on Mem0, then 50,000 adds and 10,000 retrievals a month with your own LLM key; usage past the limits billed with a monthly cap you set | Custom: limits and terms set by agreement; no published price | graph memory, user profiles, Dream, on-prem deployment, audit logs and SSO only on Custom; the earlier Starter, Growth and Pro plans are "no longer sold to new accounts" (pricing FAQ, added 26 October 2026) |
| Zep | Free: 10,000 credits a month, 2 projects, 1 Memory MCP Server seat | Flex $125 a month (50,000 credits); Flex Plus $375 a month (200,000 credits); Enterprise custom | retrieval, storage, threads and users are not metered |
| Letta | Free $0, 3 stateful agents | Pro $20 a month with $20 of credits and up to 20 stateful agents; Teams Pro $20 per seat a month; Enterprise custom | server-side tool execution billed at $0.00015 a second |
| Cognee | Free $0, 1 workspace | Standard $1.00 per 1M tokens plus $5 per additional workspace; Enterprise as a BYOC engagement | the free token allowance reads 1M on the pricing page and 10M in the billing docs |
| Supermemory | Free $0 with $5 of credits a month | Pro $19; Max $100; Scale $399 (a month) | "Coding plugins are available on every plan, including Free." (billing docs) |
| Hindsight | self-hosted free | Cloud by usage: Retain $10.00 per 1M tokens, Recall $0.75 per 1M tokens, Reflect $0.05 per call, storage $0.25 per 1M tokens a month; Enterprise by contact | "There is no fixed monthly fee and no per-seat pricing." |
| MemMachine | Open Source $0 | Pro $79 a month per project, 15% less when billed yearly; Enterprise custom | |
| EverOS | open source; Cloud Free $0 (3 memory spaces, 50,000 MCU a month) | Pro $25 a month; Enterprise custom | the pricing page does not define MCU |
| TencentDB Agent Memory | open source | Tencent Cloud, pay as you go by the hour: storage 0.02 yuan per 10,000 records, model calls 0.01 yuan per credit | on cloud.tencent.com, pages dated 7 July 2026 |
| MemoryLake | Free $0, 300,000 tokens a month | Pro $19 (6,200,000 tokens); Premium $199 (66,000,000 tokens) a month | about $3.125 per 1M tokens on demand |
| Memori | open source, self-hosted: "Start free and open. Bring your own database." | Team from $60K a year; Business from $150K a year; Enterprise custom | production plans include "Unlimited memories created & recalled" |
| Redis Agent Memory | no Agent Memory price on the pages read | a Redis Cloud database is required: Free up to 30 MB; Essentials from $0.007 an hour; Pro from $0.014 an hour with a $200 monthly minimum | |
| MinnsDB | open source (AGPL-3.0); platform plans: 30-day free trial | Hobby £20; Pro £199; Scale £499 (a month); Enterprise custom | platform plans in a collapsed section of minns.ai/pricing |
| MemNexus | Free $0, 1,000 memories held | Pro $25 a month; Enterprise custom, "Starting at $49/seat/month for teams of 5+" | |
| Sentra | Team shown as Free under a "Launch Special", with $16 per user a month struck through | Enterprise custom | the comparison table ticks "API/MCP integration" only for Enterprise, while llms.txt says "To get started, book a demo at https://www.sentra.app/demo." and an API key comes after a workspace is set up (read 9 Nov 2026) |
| LangMem | open-source library | no pricing page | |
| Mnemoverse | Free $0: 1,000 queries a day, 10,000 atoms | Pro $29 a month; Team $149 a month; Enterprise custom |
Funding, as each company states it
| Company | What its pages or the cited press say |
|---|---|
| Mem0 | "our $24M raise across Seed and Series A" (post of 28 October 2025): Seed led by Kindred Ventures, Series A by Basis Set Ventures, with Peak XV Partners, GitHub Fund and Y Combinator |
| Zep | no round, amount or date on its site; the About page names backers by logo: Root Ventures, Y Combinator and Engineering Capital, plus angels at Vercel, Google and Airtable |
| Letta | names backers on its home page, including Felicis Ventures, Essence VC and Sunflower Capital, with no round, amount or date |
| Cognee | a $7.5M seed on 19 February 2026, led by Pebblebed with 42CAP and Vermilion, and $1.5M on 18 November 2024 (read 5 October 2026) |
| Supermemory | a $3 million first round on 6 October 2025, led by Susa Ventures, Browder Capital and SF1.vc |
| Vectorize (Hindsight) | a $3.6M seed round led by True Ventures, April 2024 |
| MemMachine | no round stated; its pricing FAQ says MemMachine was incubated by MemVerge |
Read 8 October 2026 unless noted. Funding for the systems that entered the report on 13 October was not read.
Notes on individual systems
Mem0. Memory Decay is an opt-in, per-project search-time bias, off by default, that scales a memory's ranking score between 0.3x and 1.5x by recency of use; the Dream docs page says Supersede and Merge run on all plans as memories are added and Synthesis is opt-in on Pro and above, while the pricing page read on 9 November 2026 lists only Free and Custom and puts Dream under Custom, so the two pages disagree on that date; expiry is documented as expiration_date, after which a memory is hidden, not deleted (249 docs pages, 8 October 2026).
Zep. Facts stay and are invalidated rather than deleted. On top of facts Zep derives Observations (Flex Plus and Enterprise, announced 21 May 2026), cross-fact patterns that are "deduplicated and merged" as new evidence arrives. Zep's engineering blog of 27 August 2026 says retention policies expire data on a customer-defined schedule, with the graph as the unit of retention (read 8 October 2026).
Letta. Outdated memory is removed or rewritten by the agent itself, and memory changes are committed to git. The repository letta-ai/letta is now a stub; the source moved to letta-ai/letta-code (read 8 October 2026).
Cognee. Since v1.6.3 (7 October 2026) Cognee documents SearchType.TEMPORAL reranking of hybrid candidates by date fit, next to opt-in, experimental truth-subspace reranking. The documented expiry is for the session tier (default 7 days) and the documented decay is for per-user preference weights (read 8 October 2026).
Supermemory. Its v3 and v4 APIs are deprecated and will be shut down on December 31, 2026, per the docs banner. The OpenClaw plugin's README still says "Requires Supermemory Pro or above", while the billing docs and the changelog of 15 August 2026 say coding plugins are on every plan (read 8 and 13 October 2026).
Hindsight (Vectorize). Memories carry occurrence and mention times and configurable recency weighting, and no TTL or validity window; read-time deduplication of observations is documented (prefer_observations). Audit logging is documented in the server configuration and is off by default (read 8 October 2026).
MemMachine. v0.4.0 was released on 8 October 2026 with BEAM support, a LiteLLM provider, and Milvus and Qdrant backends. No TTL, expiry, supersession, validity field or contradiction rule is documented (read 8 October 2026).
EverOS (EverMind). The open-source runtime and the cloud service share the name; the runtime is Apache-2.0 and the cloud plans are listed above. No MMR or diversity setting for retrieval is documented on its 104 docs pages (read 8 October 2026).
LangMem. LangMem's own pages document no TTL; LangGraph's store takes one, per the TTL configuration page on docs.langchain.com ("If omitted, items do not expire by default", read 9 November 2026). The PyPI release 0.0.30 is dated 27 October 2025 (read 8 October 2026).
TencentDB Agent Memory. The repository TencentCloud/TencentDB-Agent-Memory holds three product lines on three branches: the default branch feat/server_team is the v2 team hub (release v2.0.2 on 9 October 2026), feat/server is the v1 line, and main is v0.3.6, so any comparison has to name the branch or version; this report describes the default branch. At extraction an LLM decides to store, skip, update or merge, with extraction prompts per layer that can be set per instance, team or agent, and agents can be excluded from capture. Skills are versioned and the previous version stays. Clean-up of older layer files by age exists and is off by default. The cloud edition was found on the Chinese site; on the international site it was not found at the two addresses tried (read 9 and 10 October 2026).
MemoryLake. Conflicting facts go to a reviewer rather than being overwritten ("MemoryLake doesn't silently keep both or overwrite blindly."), and forgetting a fact keeps its history. A product page says "Roll back, branch, and merge your memory timeline with confidence."; rollback, branching and merging of memory are not described on the documentation pages read on 13 October 2026.
Memori (Memori Labs). The vendor's engine extracts the memories ("Background AI processing extracts facts, preferences, and relationships from your conversations automatically."), and a write needs attribution ("Without attribution, Memori cannot create or recall memories."); on the 84 docs pages read no configurable write gate is described. Deleting all memories of an entity exists in the SDK only in the bring-your-own-database mode. Memori Labs is not the memoripy package, which is a different project (read 13 October 2026).
Redis Agent Memory. The open-source Agent Memory Server moved into the repository's V0/ folder on 9 July 2026, and the maintained product is the managed Redis Agent Memory in Redis Iris. In Iris, a model decides whether a near-identical memory is a true duplicate, and detectors can redact or drop sensitive data at extraction (read 13 October 2026).
MinnsDB. The vendor calls it both a memory layer ("Minns Memory Layer is a purpose-built database for agentic workflows.") and "The agentic database", so it sits on the border of the database exclusion; it extracts facts from conversations itself, which is why it counts. Every edge carries valid_from and valid_until, and the README marks memory consolidation and tiers as work in progress. The latest tag is v2.8.2 (31 May 2026), and main was updated on 5 October 2026 (read 13 October 2026).
MemNexus. A closed-source hosted service whose npm client packages declare MIT. Named memories keep versions through a SUPERSEDES chain, and search excludes superseded and contradicted memories by default. On team memory the vendor's pages say three things: a team workflow guide says "Everyone on the team uses the same API key (or keys scoped to the same workspace). When you save a memory, it's visible to everyone.", and a memory's owner can share it with named users or the whole organization; the roadmap lists shared memory spaces for teams as "Planned"; and the Enterprise plan lists "Shared team memory" (read 13 October 2026).
Sentra. Sentra describes itself as "the managed organization memory, the company brain a company's teams and AI agents share", filled from a company's connected tools and delivered to agents over REST and MCP. Its own comparison points a reader who needs memory for a single agent to other systems: "If your question is "which memory tool for my agent," pick from the six." Its llms.txt says the public API is read-oriented and ingestion happens through connected source systems, not a public write endpoint, while its comparison article describes "one governed graph of time-stamped facts that every agent and every person reads and writes" (read 13 October 2026).
Benchmark numbers as claimed by vendors
Every number in the table below is a vendor's own, taken from the vendor's page, paper, repository or benchmark, with the recipe fields that same source states; "not stated" means the cited source does not give the field. The table is here to show why these numbers do not line up with one another. Each row answers a different question under a different recipe, so the column of figures is not a ranking.
| System | Benchmark | Vendor's figure | Questions | How much memory the answering model sees (top-k) | Judge | Answering model | Who measured | Date |
|---|---|---|---|---|---|---|---|---|
| Mem0 | LoCoMo | 92.5% (1,425 of 1,540), the README headline | 1,540 | top 200 | not named in the README (the harness default is gpt-4o) | not named in the README | Mem0 | README at commit 4b61c5d, 13 May 2026 |
| Mem0 | LoCoMo | 91.56% (1,410 of 1,540), the committed results file | 1,540, categories 1 to 4 | top 200 | gpt-5 | gpt-5 | Mem0 | an earlier run: results file stamped 6 April 2026, committed 9 April 2026 (edcd6f1) |
| Zep | LoCoMo | 94.7% (1,459 of 1,540) | 1,540, four categories (the page's per-category counts add up to 1,436 of 1,539) | "Multi-scope retrieval depth: 20 edges, 10 nodes, 10 episodes, 5 thread summaries, 5 observations, cross-encoder reranking"; median context 5,760 tokens | gpt-5.4 "with chain-of-thought grading" | gpt-5.4 (reasoning = medium) | Zep | no run date published; research page dated 28 May 2026 in its earliest archived copy (14 Jun 2026), same figures and recipe, last modified 29 Sep 2026; since November 2026 on its own page, research/memory-benchmarks, same figures and recipe (read 9 Nov 2026) |
| Zep | LongMemEval, variant not stated | 90.2% (451 of 500) | 500 | as in the row above; median context 4,408 tokens | gpt-5.4 "with chain-of-thought grading" | gpt-5.4 (reasoning = medium) | Zep | no run date published; research page dated 28 May 2026 in its earliest archived copy (14 Jun 2026), same figures and recipe, last modified 29 Sep 2026; since November 2026 on its own page, research/memory-benchmarks, same figures and recipe (read 9 Nov 2026) |
| Supermemory | LongMemEval-S | 97%, labelled "Recall@20 with aggregation" | 500 | not stated as a number of memories (the report gives "The retrieval configuration is Recall@20 with aggregation") | gpt-4o | gpt-4o | Supermemory | no run date published; report page read 8 and 17 Oct and 9 Nov 2026 |
| Hindsight | LongMemEval-S | 91.4% | 500 | not stated | GPT-OSS-120B, temperature 0 | Gemini-3 Pro | Vectorize, in its own paper | arXiv 2512.12818 v1, 14 Dec 2025 |
| Hindsight, on AMB | LongMemEval-S | 0.946 (473 of 500) | 500 | not stated as a number of memories | gemini-2.5-flash-lite | gemini-3.1-pro-preview | Vectorize, which builds and runs AMB | no run date published |
| Hindsight, on AMB | LoCoMo10 | 0.920 (1,417 of 1,540) | 1,540 | not stated as a number of memories | gemini-2.5-flash-lite | gemini-3.1-pro-preview | Vectorize | no run date published |
| Hindsight, on AMB | BEAM 10M | 0.6408, a mean score, not a share of correct answers | 200 | not stated as a number of memories | gemini-2.5-flash-lite | gemini-3.1-pro-preview | Vectorize | no run date published |
| EverOS | LoCoMo | 94.42% | 1,540 | top_k 20 (episodes per question) | gpt-4o-mini, one judgement per question | gpt-4.1-mini | EverMind | configuration at commit c479f45, 8 Sep 2026 |
| TencentDB Agent Memory | SWE-bench, the 50 hardest tasks run in one session | pass rate 17% without team memory, 20% with it | 50 tasks | not stated | the tasks' own tests, no LLM judge | not stated | Tencent | vendor article, 18 Aug 2026 |
What the rows leave out, by row:
- Mem0. The two Mem0 rows come from two different runs. The README names no judge or answering model for its headline; the committed results file records the earlier run, with gpt-5 in both roles. The two rows belong together, and the results file's recipe is not the recipe of the README headline.
- Zep. Both figures come from Zep's research page, not from its arXiv paper 2501.13956, which reports DMR 94.8% and LongMemEval_s 71.2%. The page's per-category LoCoMo counts add up to 1,436 of 1,539, not to the headline's 1,459 of 1,540.
- Supermemory. The figure is a recall measure, not a share of answers judged correct, and its label is quoted as the vendor gives it. The page gives no date for the figure (in October it sat under a section heading "May 2026", which the page no longer carries on 9 November 2026); 97% at Recall@20 first appears in an archived copy of 1 October 2026, and the same table earlier gave other numbers for gpt-4o. The repository README still gives 95% at Recall@15.
- Hindsight on AMB. AMB is the Agent Memory Benchmark that Vectorize, the company behind Hindsight, built and runs; its run files carry no dates. The BEAM 10M figure is a mean score, not a share of questions answered correctly.
- EverOS. Memories are extracted by gpt-4.1-mini. The vendor names the decider model two ways: deepseek-v4-flash-0731 in the README table, and qwen3.6-27B in a comment of the configuration file, which says the "published numbers used qwen3.6-27B" (README and configuration at commit c479f45, read 17 October 2026).
- TencentDB Agent Memory. The article names neither the model nor the number of runs, and the vendor writes that before the conclusion is widened the task subset has to be fixed and the sample size and confidence intervals reported.
Redis publishes a LongMemEval-S figure, 86.5% over 500 questions with gpt-4o answering and judging (report of 21 September 2026), for a research configuration, and the README of its open agent-memory-server repository says of it: "Do not treat 86.5% as a Redis Cloud product score." It is a mean over six question types ("task-averaged accuracy"), not a share of all questions answered correctly, so it is kept out of the table rather than set beside the item scores there. A Redis AI Research note labelled June 2026, online since 28 July 2026, gives 86.1% for the same combination.
Why these numbers do not compare with each other, read straight from the table:
- Questions. From 50 tasks to 200, 500 and 1,540 questions. Every LoCoMo row counts 1,540; Mem0's results file records categories 1 to 4, and EverOS's configuration says the adversarial category 5 is dropped.
- Memory shown to the answering model. From 20 episodes to 200 memories to a typed bundle of edges, nodes, episodes, summaries and observations, or not stated.
- Judges. gpt-5, gpt-5.4, gpt-4o, gpt-4o-mini, GPT-OSS-120B and gemini-2.5-flash-lite, the tasks' own tests in one row, and no judge named in one.
- Answering models. Six different models, and none named in two rows.
- Metrics. A share of answers judged correct, a mean score, a recall label and a task pass rate; Redis's figure below the table is an average across question types.
- Who measured. Every row is a vendor measuring its own system, or a benchmark that a vendor builds and runs.
- One vendor, two runs. Mem0's README and its own results file give two different LoCoMo figures.
Further vendor figures are not in the table because the cited source does not state a field the table needs (sources read 8 to 13 October 2026): Mem0's LongMemEval figure (the README names no judge, and no results file for it was read), Letta's LoCoMo result (no judge named in the post), Cognee's BEAM results (judge, answering model and depth not named), MemMachine's LoCoMo results (judge not named), the LoCoMo figure in Hindsight's paper (the number of questions is given as "Varies"), MemoryLake's LoCoMo figure (the metric is described three different ways and the models are not named), Memori's LoCoMo figure (the number of questions is not stated), TencentDB's other runs, whose recipes were not fully established from the pages read, and Sentra's Terminal-Bench run, which tests Sentra Code Memory, a product not yet released, is pending verification by the benchmark's team, and is not a memory benchmark. Systems named in one line above carry no numbers here.
Memory built into assistant platforms
Platform memory is outside the selection rule, because it comes with the assistant rather than being connected by a developer. Products read: ChatGPT, Codex and Computer History (OpenAI); the Claude apps, Claude Code, the memory tool and Managed Agents memory stores (Anthropic); the Gemini app and Gemini CLI (Google); Copilot for individuals, Microsoft 365 Copilot, Copilot Studio and GitHub Copilot Memory (Microsoft); the Grok apps, Grok Skills, Grok Bot and Grok Build (xAI). Google's Memory Bank and Microsoft Foundry were not read, so nothing here speaks for them.
| Platform | Memory launch, in the vendor's own source | Export, as documented |
|---|---|---|
| OpenAI ChatGPT | announced 13 February 2024; Plus users 29 April 2024; Free 5 September 2024; chat-history referencing 10 April 2025 | the account export ZIP "includes your chat history and other relevant account data", and that sentence does not name memory |
| OpenAI Codex | local memories, off by default, first in the changelog section dated 16 April 2026; memory files are plain Markdown under ~/.codex/memories/ | files on disk |
| Anthropic Claude apps | Team and Enterprise 11 September 2025; Pro and Max 23 October 2025; Free 2 March 2026 | memory import and export documented |
| Anthropic, developer surfaces | memory tool in beta 29 September 2025; Claude Code auto memory in 2.1.59, 26 February 2026; Managed Agents memory in public beta 23 April 2026 | Managed Agents memories can be read, edited, imported and exported through the API or the Console |
| Google Gemini app | past chats for Advanced subscribers 13 March 2025; for all users 13 August 2025; "Past chats" renamed "memory" 26 March 2026 | Google Takeout exports Gemini Apps data, and memory is not named in it |
| Microsoft | Copilot for individuals 4 April 2025; Microsoft 365 Copilot memory 14 July 2025; GitHub Copilot Memory in public preview 15 January 2026 | "Export all activity history" gives a CSV, with memory not named; Microsoft 365 admins can export memory data through eDiscovery and Graph; GitHub admins can export Copilot memories |
| xAI Grok | no xAI page read dates the launch of memory in the Grok apps; Grok Skills 18 May 2026; Grok Bot 11 August 2026; Grok Build memory 16 September 2026 |
Anthropic and Microsoft dates were read on 5 October 2026, OpenAI and Google dates on 8 October. A few limits stated by the platforms themselves: Managed Agents memories are "capped at 100 kB (~25k tokens)" each, "A store holds a maximum of 10,000 memories.", and "A maximum of 8 memory stores is supported for each session." (wording read 9 November 2026); Claude Code loads the first 200 lines or 25KB of MEMORY.md; and GitHub Copilot Memory, read 9 November 2026, says stored facts and preferences "expire after a period without reinforcement" and "The default expiry period is 90 days.", while Copilot's own votes decide what is used: "Upvoted memories are more likely to be sent." and "Memories that are downvoted past a threshold are no longer used and will be deleted after a period of disuse." In Codex, ChatGPT web uses ChatGPT memory while local Codex clients use a separate local memory store; Codex documents an --oss mode with a local provider, and whether memory works in it is not stated. Perplexity Brain, a memory system for Perplexity Computer, rolled out on 18 June 2026 to Max and Enterprise Max subscribers in Research Preview.
Is there an agent memory standard?
No agent memory interoperability standard has been ratified by the IETF, the W3C, the Linux Foundation (AAIF, LF AI & Data) or the MCP project, on the pages read on 7 and 13 October 2026.
- W3C. The AI Agent Memory Interoperability group is a W3C Community Group, which under the W3C process does not produce W3C standards ("Per the W3C process, a Community Group does not produce W3C Recommendations or other W3C standards."). Its charter v1.0 was adopted on 19 June 2026; it has 33 participants representing 10 organizations and 19 messages in its list archive from 5 June to 12 October 2026 (read 13 October 2026).
- IETF. The group's chair posted on 25 July 2026 that the IETF Independent Submission of draft-saihm-memory-protocol had concluded, and on the datatracker that draft is in no RFC stream. Other individual Internet-Drafts on agent memory include a memory interchange bundle format (draft-vu-aimem-bundle, 14 June 2026), an architecture and data model for persistent agent memory (draft-infantado-agent-memory-architecture, 19 July 2026) and a benchmarking method for the integrity of agent memory at rest (draft-khandelwal-bmwg-agent-memory-integrity, 2 October 2026). The proposed agentproto working group's draft charter leaves conversation memory out: "It does not include the content exchanged with AI models, such as conversation memory, retrieved documents, or prompts." (read 7 October 2026)
- MCP and A2A. The MCP specification of 28 July 2026 mentions a memory tool once, as an example ("For example, the world of a web search tool is open, whereas that of a memory tool is not."). The A2A specification describes agents working together "without needing access to each other's internal state, memory, or tools" (read 7 October 2026).
What a memory interoperability layer would need to specify is laid out in The Missing Layer: No Protocol Says What Agents Know.
What changed since April 2026
Dated changes found on the pages read, oldest first:
- 23 April 2026: Claude Managed Agents memory entered public beta.
- 21 May 2026: Zep announced Observations, cross-fact patterns on Flex Plus and Enterprise.
- 3 June 2026: the W3C AI Agent Memory Interoperability Community Group was created, and Claude Code 2.1.162 renamed Windsurf to Devin Desktop in its
/idemenu. - 18 June 2026: Perplexity Brain rolled out in Research Preview.
- 19 June 2026: the W3C group adopted its charter v1.0.
- 9 July 2026: Redis moved its open Agent Memory Server into
V0/; the managed Redis Agent Memory in Redis Iris is in preview. - 23 July 2026: Cognee's VS Code extension, version 0.1.0, was last updated on the Visual Studio Marketplace, and the IETF agentproto BoF met at IETF 126.
- 13 August 2026: Mem0's shared editor plugin added Kimi Code on the same memory bank as Claude Code, Cursor, Codex and Antigravity.
- 15 August 2026: Supermemory said no plugin sits behind a paid plan any more.
- 18 August 2026: Tencent published the team-memory article behind its SWE-bench figures.
- 16 September 2026: Grok Build memory became available.
- 4 to 6 October 2026: Mnemoverse's REST read API took a diversity setting on 4 October, and its MCP package 0.15.0 and Python SDK 0.5.0 on 6 October.
- 7 October 2026: Cognee v1.6.3 added TEMPORAL reranking.
- 8 October 2026: MemMachine v0.4.0.
- 9 October 2026: TencentDB Agent Memory v2.0.2, the team hub line.
- 26 October 2026: Mem0's pricing moved to two plans, Free and Custom; its new pricing FAQ says the earlier paid plans are "no longer sold to new accounts".
Graphlit, which some earlier comparisons list, is winding down and no longer takes new customers.
Mnemoverse, the publisher's system
We build Mnemoverse, a hosted memory API for AI agents. It is reached through REST, a Python SDK, the @mnemoverse/mcp-memory-server package over stdio (version 0.17.1 on npm, 9 November 2026) and a hosted MCP connector. What our documentation states:
- Outcome input. REST "Report outcome (success/failure) for memories", the MCP tool
memory_feedbackand the SDK'sfeedback. Ratings re-rank what comes back next. - Write gate. A write the embedder cannot tell apart from one already stored in the same domain is not stored.
- Diversity on read. Since 4 October 2026 the REST read API takes a diversity setting (0 to 1, default 0; above 0 the cut becomes a maximal marginal relevance selection); memory_read in the MCP package 0.15.0 and read() in the Python SDK 0.5.0 take it since 6 October 2026.
- Replacing a memory. A write can name the memories it replaces through REST and the Python SDK; no MCP tool accepts that field today. A replaced memory stays fetchable by id.
- Consolidation. "Consolidation is off on the hosted service and we do not currently recommend turning it on." The endpoint returns 503 with a typed reason (API reference).
- Across clients and people. "Any key or sign-in from the same Mnemoverse account reaches the same memory." (Coding assistants) Rooms (beta) share memory across accounts.
- Hosts. Per Install by host on 17 October 2026, Shipped: Claude Code (plugin), Claude Desktop, Cursor (plugin), VS Code with Copilot Chat (extension), Gemini CLI, Kimi Code (manifest), Devin Desktop, formerly Windsurf (
npxpackage), and Cline, Zed and any MCP client; Verified live: Hermes Agent and OpenClaw, through a server entry. - Keys. The VS Code extension by default runs
npxlocally with a key that browser sign-in creates and the OS keychain stores, so nothing is pasted, and its hosted mode needs no key. The remote server and the Claude Code, Cursor and Gemini CLI plugins sign in through the browser. An API key goes into the Claude Desktop bundle (.mcpb) and into thenpxpackage entry used for Devin Desktop and other MCP clients. - Self-hosting. Enterprise customers can self-host by agreement; no self-hostable build is published.
- Price. Free $0 (1,000 queries a day, 10,000 atoms), Pro $29 a month, Team $149 a month, Enterprise custom (pricing, read 13 October 2026).
We publish how often our own feedback endpoint is used. From Agent memory feedback: the missing signal: "In the campaign pass of 2026-08-18, 33 of the 62 live tenants never called the feedback endpoint and the median share of edges touched by feedback is 0.0% (highest 28.3%); benchmark-grown memory, from our own harness, had 94.8%."
To try it with your own agent, start from Getting Started.
Common questions
How do available AI agent memory layers compare for developer workflows?
On the questions that decide a developer workflow, read in October 2026: a documented call that marks a specific recalled memory helpful or wrong exists at Mem0 (hosted platform only), MinnsDB and Mnemoverse; Hindsight marks a stored fact invalid; Cognee and Supermemory route answer ratings and reviews of inferred memories into retrieval; MemNexus records which search results were selected; and Memori takes free-text feedback with no memory id. All six systems of the comparison page Best Persistent Memory APIs for AI Agents document one memory reached from several clients. Mem0, Letta, Cognee and Supermemory document a way to run without a vendor account, and Zep does through Graphiti; Mnemoverse does not. Vendor benchmark scores come from different recipes and do not rank the systems.
What is the state of AI agent memory in 2026?
In October 2026, 23 products qualify as agent memory layers under this report's selection rule, from the six systems of the comparison page Best Persistent Memory APIs for AI Agents (Mem0, Letta, Zep with Graphiti, Cognee, Supermemory and Mnemoverse) to other layers such as Hindsight, MemoryLake, TencentDB Agent Memory and MinnsDB. ChatGPT, Claude, Gemini, Microsoft Copilot and Grok each ship memory of their own. No ratified interoperability standard exists, and benchmark headlines are measured by vendors under recipes that differ in judge, answering model and memory depth.
Which AI agent memory layers learn from whether a recalled memory helped?
Counting a documented API, SDK method or MCP tool that marks a specific recalled memory helpful or wrong after use: Mem0 has one on its hosted platform, MinnsDB has one as two REST events, and Mnemoverse has one in REST, MCP and its SDK. Hindsight documents marking a single stored fact invalid; Cognee and Supermemory document a path into retrieval from answer ratings and from review of inferred memories; MemNexus records whether a search result was selected, for measurement; Memori takes free-text feedback without a memory id; Letta labels agent steps. On the pages read, Zep, MemMachine, LangMem, EverOS (outside a pre-release cloud SDK that records a rating without applying it), TencentDB Agent Memory, MemoryLake, Redis Agent Memory and Sentra document no such input.
Which agent memory systems can I self-host without a vendor account?
Among the six systems of the comparison page: Mem0 Open Source, Letta's local runtime, Cognee's open-source engine, Supermemory's closed self-hosted binary, and Zep only through Graphiti; for Mnemoverse, Enterprise customers can self-host by agreement; no self-hostable build is published. Beyond them, Hindsight and TencentDB Agent Memory (MIT), MemMachine and EverOS (Apache-2.0), MinnsDB (AGPL-3.0, with a Qdrant store), the LangMem library and Redis's open V0 server, which Redis no longer maintains, run without an account; Memori stores memory in your own database but calls its vendor's API for extraction. MemoryLake, MemNexus, Sentra and Redis Agent Memory in Redis Iris show no account-free path on the pages read.
Can agent memory benchmark scores be compared across vendors?
Not directly. Published headlines differ in the question set, in how much memory the answering model sees, in the judge and in the answering model, and each is measured by the vendor itself or on a benchmark a vendor builds and runs. A score compares only with another produced under the same recipe, which is why this report prints vendor numbers in one table with their recipes rather than as a ranking.
Is there a standard for AI agent memory?
No ratified one as of October 2026. The W3C AI Agent Memory Interoperability Community Group adopted its charter on 19 June 2026, and under the W3C process a Community Group does not produce W3C standards. The IETF has individual Internet-Drafts on agent memory and a proposed agentproto working group whose draft charter leaves conversation memory out.
Which memory layer works with coding agents such as Claude Code, Cursor and Codex?
Most of them, through an MCP server or a plugin, and each coding tool still gets its own setup. Mem0 keeps plugin directories for ten agent hosts, and its editor plugins share one memory bank; Supermemory maintains seven host plugin repositories; Zep ships manifests for Claude, Codex and ChatGPT Work, Cursor and the Agent Plugins format, and sends VS Code, Copilot and Gemini CLI to its Context MCP Server; Cognee ships manifests for Claude Code, Codex and Antigravity and a VS Code extension; Letta reaches Cursor and VS Code through its hosted MCP server; Hindsight's Coding Agents integration listed 22 harnesses on 8 October 2026; MemNexus's CLI sets up seven clients; Mnemoverse ships a Claude Code plugin, a Cursor plugin, a VS Code extension and a Gemini CLI extension.
Where Mnemoverse comes out worst
- No account-free path. Of the six systems on the comparison page Best Persistent Memory APIs for AI Agents, Mnemoverse is the one without a way to run it with no vendor account. The engine is closed; Enterprise customers can self-host by agreement; no self-hostable build is published. Mem0, Letta, Cognee and Supermemory document such a path, and Zep does through Graphiti.
- Memory lives on our servers. The npm package runs locally over stdio but calls the hosted API, which rules it out where memory has to stay on the developer's own machine.
- Consolidation has never completed. In our API reference's words, "It has never once completed on the hosted service"; it is switched off and the endpoint returns 503.
- The feedback channel is barely used. In the census of 18 August 2026 quoted in the section above, "33 of the 62 live tenants never called the feedback endpoint".
- Some routes still take a pasted key. The Claude Desktop bundle (
.mcpb) and thenpxpackage entry for Devin Desktop and other MCP clients use an API key, while the remote server, the Claude Code, Cursor and Gemini CLI plugins and the VS Code extension sign in through the browser.
Cite this report
Izgorodin, E. (2026). State of AI Agent Memory 2026: How do available AI agent memory layers compare for developer workflows? Mnemoverse. https://mnemoverse.com/docs/research/state-of-agent-memory-2026. Facts read 5 to 13 October 2026 and re-read on 9 November 2026.
One line to quote: "An agent memory benchmark score is the output of its recipe (the question set, the memory shown to the answering model, the judge and the answering model), so two scores produced under different recipes do not rank the systems behind them."
Sources
Read between 5 and 13 October 2026, Mnemoverse's own hosts and two configuration files on 17 October, each with two instruments and a nonexistent address on the same host as a 404 control. Every quotation and every price was re-read on 9 November 2026; what had changed by then is rewritten and dated in the text. Dates per claim are in the text.
Mnemoverse
- API reference, Getting Started, Install by host, VS Code, Rooms (beta), API pricing, pricing, LoCoMo recipe page, Agent memory feedback: the missing signal
Mem0
- docs.mem0.ai, changelog highlights, platform vs open source, pricing, pricing FAQ, Dream, memory-benchmarks README at 4b61c5d, results file at 4b61c5d
Zep and Graphiti
- help.getzep.com, February 2026 deprecation wave, Context MCP Server, observations, FAQ, memory benchmarks, pricing, zep-context-plugin
Letta
Cognee
Supermemory
- docs, memory review, billing, changelog of 15 August 2026, LongMemEval report, pricing, OpenClaw plugin README
Hindsight (Vectorize)
- hindsight.vectorize.io, pricing, about, arXiv 2512.12818, Agent Memory Benchmark, vectorize-io/hindsight
MemMachine, EverOS and LangMem
- memmachine.ai pricing, EverOS benchmarks README, EverOS LoCoMo configuration, EverOS pricing, LangMem, LangGraph store TTL
TencentDB Agent Memory
MemoryLake, Memori, Redis Agent Memory
- MemoryLake docs, MemoryLake pricing, MemoryLake changelog
- Memori docs, Memori pricing, MemoriLabs/Memori
- Redis Agent Memory overview, self-managed, Redis Cloud pricing, Google ADK integration, benchmark report, redis/agent-memory-server
MinnsDB, MemNexus, Sentra
- docs.minns.ai, event types, Minns-ai/MinnsDB, pricing
- MemNexus docs, team workflows, memories API, pricing
- Sentra llms.txt, comparison article, MCP access, pricing
Platforms
- OpenAI: memory and new controls, Codex memories
- Anthropic: release notes, Managed Agents memory, Claude Code memory
- Google: past chats for all users, rename to memory
- Microsoft: Copilot activity history, Copilot memory, GitHub Copilot Memory
- xAI: Grok Skills, Grok Bot, Grok Build memory
- Perplexity: What is Brain
Standards
- W3C AI Agent Memory Interoperability Community Group, its list archive, IETF agentproto, draft-vu-aimem-bundle, draft-infantado-agent-memory-architecture, draft-khandelwal-bmwg-agent-memory-integrity, Cursor plugins
The 28 third-party pages behind the selection rule
- https://mem0.ai/
- https://powerdrill.ai/blog/best-ai-agent-memory-solutions
- https://atlan.com/know/best-ai-agent-memory-frameworks-2026/
- https://vectorize.io/articles/best-ai-agent-memory-systems
- https://www.braintrust.dev/articles/best-ai-agent-memory-tools-2026
- https://evermind.ai/blogs/8-best-ai-agent-memory-frameworks-for-developers-in-2026
- https://www.cognee.ai/best-ai-memory-layers-for-ai-agents-in-2026-comparison
- https://dev.to/jonathanfarrow/the-10-best-ai-memory-layers-for-agents-in-2026-448e
- https://www.cognee.ai/best-open-source-ai-memory-tools-for-llm-agents-and-developers
- https://www.stork.ai/blog/best-memory-layer-ai-agents-2026
- https://www.graphlit.com/blog/survey-of-ai-agent-memory-frameworks
- https://dev.to/thedailyagent/top-6-ai-agent-memory-frameworks-for-devs-2026-1fef
- https://omidsaffari.com/blog/best-persistent-memory-systems-for-ai-agents-2026
- https://aiworkflowlab.dev/article/agent-memory-mem0-vs-letta-vs-zep-2026
- https://www.contextstudios.ai/guides/best-ai-agent-memory-tools-2026
- https://www.developersdigest.tech/blog/best-ai-agent-memory-providers-2026
- https://techsy.io/en/blog/best-ai-agent-memory-tools
- https://coworker.ai/blog/mem0-vs-zep-vs-letta
- https://www.puppyone.ai/en/blog/best-ai-agent-memory-platforms
- https://supermemory.ai/blog/best-memory-apis-stateful-ai-agents/
- https://www.shaped.ai/blog/the-8-best-tools-for-ai-agent-memory-long-term-recall-2026-guide
- https://forum.letta.com/t/agent-memory-letta-vs-mem0-vs-zep-vs-cognee/88
- https://greennode.ai/blog/memory-architecture-for-ai-agents
- https://mem0.ai/blog/vector-databases-and-memory-for-ai-agents
- https://aiopsschool.com/blog/top-10-agent-memory-stores-features-pros-cons-comparison/
- https://mem0.ai/blog/mem0-vs-zep
- https://memnexus.ai/blog/2026-07-23-best-ai-agent-memory-systems
- https://foundrysoft.co/blog/ai-agent-memory-mem0-vs-zep-vs-letta
Related
- Best Persistent Memory APIs for AI Agents: the six systems of the first selection condition, on six questions
- LoCoMo recipe page: one protocol, its graders, and dated vendor rows
- Agent memory feedback: the missing signal: how often the outcome input is used in practice
- MCP servers that share state across IDEs: how one memory reaches several clients
- The Missing Layer: No Protocol Says What Agents Know: what a memory interoperability standard would have to specify
Facts read 5 to 13 October 2026; Mnemoverse's own hosts and package version on 17 October. Corrections and additions: hello@mnemoverse.com
Mnemoverse is a persistent-memory API for AI agents. Free key: console.mnemoverse.com · Plans and limits · Docs: Getting Started