List AI agent memory layers with support for shared agent workspaces or multi-user rooms.
TL;DR. In this page's reading, a shared space, its members and the boundary a request respects are three different things: the word on a pricing page is often a billing object, and the boundary a request respects is often a narrower object, such as a namespace or a scope.
- The list. These systems document a shared space for a team, as read on 17 November 2026: Supermemory (organization, spaces, namespaces), Mem0 (organizations, projects), Zep (account, projects), Letta (organization, shared memory repositories), Cognee (workspace as tenant), Hindsight (organization, memory banks), LangSmith with Deep Agents (organization memory as a namespace you define; agent memory shared across a Managed Deep Agents deployment), Claude Managed Agents (workspace memory stores, in beta), Google Agent Platform Memory Bank (scopes), AWS AgentCore Memory (namespaces), Microsoft Foundry memory (scopes, in public preview), Basic Memory (team workspaces), Sentra (organization memory), Caura (agent, fleet and tenant scopes) and Cloudflare Agent Memory (shared profile, in private beta). Mnemoverse, this page's publisher, shares memory through rooms across accounts (its own documentation, read 18 November 2026).
- "Workspace" means different things. Cognee says a workspace "operates as a tenant" and "is not a database boundary", while invited users "gain access to all tenant-wide datasets"; Supermemory calls namespaces "hard boundaries"; Claude scopes memory stores to a workspace.
- Administration and plan. Invitations and seats start on different plans, and SSO is often an Enterprise feature. Zep's blog of 19 August 2026 puts its Memory MCP Server on all plans and adds: "Custom OIDC and UserGroup ABAC require the Enterprise plan. Google Workspace sign-in does not."
- After Company Brain. Supermemory stopped its hosted Company Brain on 9 September 2026; its MCP and plugins keep running, and its guide says the open-source build for your own Cloudflare account "doesn't expose an MCP server or an API for asking the brain questions".
Reading each system through five questions, what the shared space is, who manages its members, from which plan a team can share, how data leaves, and where the system runs, is this page's framing; every behaviour is taken from the vendor's own documentation, pricing page or blog on the date in Sources.
The cross-IDE comparison of 16 October covers how editors and assistants connect to shared context and who sees what by default; this page covers who runs each vendor's shared space, from which plan, how data leaves and where it runs.
Read on 17 November 2026; the Mnemoverse row, from Mnemoverse's own pages, on 18 November. "Not described on the pages read" refers to that vendor's pages under Sources, not to a product capability.
| System | Shared space | Members and roles | Sign-in, SSO, SCIM | Team sharing from (plan, date read) | Export and moving | Where it runs | Status on the read date |
|---|---|---|---|---|---|---|---|
| Supermemory | Organization, spaces, namespaces | Owner, Admin, Member; Restricted from Scale | SSO on Enterprise | Pro, $19 a month, 3 seats | JSON export | Managed cloud; Local; own cloud from Scale | Hosted Company Brain discontinued 9 September 2026 |
| Mem0 | Organizations, projects | READER, OWNER | Custom plan lists SSO | Free and Custom; member plan not described on the pages read | Export jobs | Platform; on-prem on Custom | Not described on the pages read |
| Zep | Account, projects | Account Owner and Account Admin on any plan; other roles need Enterprise | Dashboard SSO, operator-approved; SCIM on Enterprise | Free: 1 Context MCP Server seat; Flex, $125 a month: 5 seats | Not described on the pages read | Managed cloud, BYOC | SCIM in beta |
| Letta | Organization, shared memory repositories | Editor, Admin, Analyst | SAML/OIDC SSO on Enterprise | Teams Pro, $20 per seat a month | CLI export of one agent's memory | Letta Cloud, where shared repositories work; self-hosted agents | Not described on the pages read |
| Cognee | Workspace as tenant | Owner, Member | Enterprise BYOC: SSO optional (12 months) or included (24 months) | Free workspace, unlimited users | Dataset Markdown report | Cognee Cloud; Enterprise BYOC | Roles not yet populated in the UI |
| Hindsight | Organization, memory banks | Owner, Admin, Member | SSO on Enterprise; can be enabled for a shared-Cloud organization, not self-serve | Cloud, usage pricing | Whole-bank ZIP | Shared Cloud; Enterprise private cloud, BYOC, on-premises | Not described on the pages read |
| LangSmith with Deep Agents | Deep Agents: organization memory as a namespace you define; Managed Deep Agents: agent memory shared by every caller | Organization roles on Plus; RBAC on Enterprise | Plus: Google, GitHub; Enterprise: SAML SSO, SCIM | Plus, $39 per seat a month | Not described on the pages read | Cloud; Enterprise: hybrid or self-hosted | Managed Deep Agents in public beta |
| Claude Managed Agents | Workspace-scoped memory stores | Workspace roles | Console: not read for this page; self-hosted: your identity provider | Tokens plus $0.08 per session-hour | API export; past versions may be deleted after 30 days | Anthropic; self-hosted on request | Beta; self-hosted on request |
| Google Agent Platform Memory Bank | Scopes | IAM roles | Not described on the pages read | Usage: storage $0.30 per GiB-month, plus operations; model tokens billed separately | Not described on the pages read | Google Cloud, fully managed | Current billing since 1 September 2026 |
| AWS AgentCore Memory | Namespaces | IAM roles | Not described on the pages read | Usage: $0.75 per 1,000 built-in long-term records a month, plus retrieval and ingestion | Record operations: list records, for example to S3, and restore them | AWS | Not described on the pages read |
| Microsoft Foundry memory | Scopes | RBAC or keys | Not described on the pages read | Azure subscription; can change in preview | Record operations; export not described on the pages read | Azure | Public preview |
| Basic Memory | Team workspaces | Viewer, Editor, Admin, Owner; User Admin on Business | SSO add-on on Business | Cloud, $19 a month per workspace, 2 people | Markdown; project ZIP | Cloud; AGPL-3.0 self-hosting | Usage plans since 26 October 2026 |
| Sentra | Organization memory | Enterprise roles | Enterprise SSO/SAML | Team "Launch Special": "Free", "$16" struck through, "per user / month" | Not described on the pages read | Cloud, isolated VPC, air-gapped on-prem | Not described on the pages read |
| Caura | Agent, fleet, tenant scopes | Not read for this page | Not read for this page | Free tier; Pro from $49 a month | Not described on the pages read; Caura says the same API runs managed and self-hosted | Managed; Apache 2.0 engine | Not described on the pages read |
| Cloudflare Agent Memory | Shared profile | Not described on the pages read | Not described on the pages read | Not billed in private beta | Promise; R2 exports planned | Cloudflare | Private beta |
| Mnemoverse (publisher) | Rooms shared across accounts | Room owner; each member read or read-write; invite codes from the agent, members managed in the console | Google, GitHub or email; Enterprise: SSO/SAML, SCIM | Any API key, Free included; usage billed to each member's own plan | Account export: JSONL, Markdown, CSV, text; import from ChatGPT, Claude Code, Pieces; console pages in beta | Managed cloud; Enterprise: your region, dedicated infrastructure, self-hosting by agreement | Live; create, invite and join are self-service |
ChatGPT shared projects are a note, not a row; see "Where ChatGPT shared projects fit".
Which agent memory layers support a shared workspace or room
A shared space is what a vendor lets several people or agents read and write: an organization, project, workspace, namespace, scope, bank, profile or room.
Team administration is the roles, identity-provider sign-in, SSO and SCIM that decide who invites and removes members.
A system is in the table when its main job, in the vendor's words, is memory for agents or for an organization; the State of AI Agent Memory 2026 compares prices and self-hosting for 17 of the 23 products it counts as memory layers.
On 9 July 2026 snowhy wrote on Ask HN: "My cofounder and I each run our own Claude Code agent locally." Later in the same post: "We're maintaining two 'company brains' and they're drifting apart." On 12 November Naman Jain, who builds Lockstep, presented it in a dev.to post as "shared team memory for Claude Code". These are requests and product claims, not measurements of the vendors in the table.
What Supermemory discontinued on 9 September 2026, and what still runs
Supermemory launched Company Brain on 24 July 2026 as "a Slack agent with shared memory for your company", and on 13 August priced it from $100 a month on Max.
Its post of 10 September 2026 says: "The supermemory company brain and Nova are discontinued as of 9th September." Everyone charged was refunded, and "Our MCP and plugins keep working".
Three lines in that post concern Nova, Supermemory's personal knowledge management platform, not Company Brain: Nova users "can get the same functionality using our MCPs"; for Nova, "Only the surface of the app is being discontinued"; and Nova's "functionality is being moved to our console."
On 25 September the changelog said: "The hosted product has been discontinued, but its harness is now available to everyone." The Apache 2.0 repository is "rebuilt to run on your own Cloudflare account", and its outside-Slack guide documents "Export as Markdown" on Home.
For teams, Supermemory's MCP docs say "Teammates collaborate within the spaces they are allowed to read or write"; its connectors include Google Drive, Gmail, Notion, OneDrive, GitHub, Granola and Web Crawler.
This page's reading: a team that wants the same memory reachable from coding agents now reads Supermemory's own docs as two paths: the MCP with spaces, or the self-hosted Slack build, whose guide says there is "no endpoint to point a coding agent at".
What "workspace" means at each vendor
At Supermemory, "namespaces are hard boundaries between users, tenants or projects". Cognee's workspace "operates as a tenant" and "is not a database boundary"; invited users "gain access to all tenant-wide datasets". Each Hindsight bank "is a separate memory for one agent, user or context." In Deep Agents, "Organization memory is typically read-only to prevent prompt injection via shared state." Google, Microsoft and AWS isolate by scope or namespace; Cloudflare nests profiles in namespaces. At Mnemoverse, the publisher, a room is "its own memory bucket, separate from every member's personal memory", shared across accounts, behind "a database-enforced membership check, not an application-layer one".
This page's reading: the word on the pricing page is not always the boundary the data respects. Membership also does not guarantee safe retrieval, so check the access boundary at retrieval, not only the invitation screen.
Who manages members: roles, sign-in, SSO and SCIM
Access is governed in two places: the people who administer a space, and the identities its agents run under. At Basic Memory, "Agents never count as members." How many roles a vendor defines is a separate question from what it lets a team share.
Zep: "Non-Enterprise accounts have only the Account Owner and Account Admin roles. The other roles require an Enterprise plan." Hindsight's team page reads two ways on roles: "Only Owners can change member roles", yet the Admin role's list includes "Change user roles between admin and member". It also labels Members "View-only access" while listing that they may "Use the Retain, Recall, and Reflect operations via API keys". This page's reading: the two lines may describe the dashboard and the API separately, so confirm the scope with Hindsight before assigning roles.
SSO is often an Enterprise feature; Basic Memory sells it as "a paid add-on on the Business plan", and Cognee's pricing table marks it "Optional" on 12-month and "Included" on 24-month Enterprise BYOC terms. LangSmith puts SAML SSO and SCIM on Enterprise; "On cloud, SCIM requires SAML SSO." Zep's dashboard SSO runs on managed cloud after an operator approves the domains, "which is where eligibility is decided", and on 17 November 2026 "SCIM provisioning is in beta. It requires the Enterprise plan. Zep turns on SCIM for each Enterprise account on request." SCIM is not in Supermemory's documentation on 17 November 2026 (153 pages of its documentation export read). Mnemoverse members sign in with Google, GitHub or email, and Mnemoverse lists SSO/SAML and SCIM on its Enterprise plan.
Hindsight, on a departing member: "Removed members immediately lose access to the organization." At Mnemoverse, "Revocation is immediate and forward-only."
From which plan a team can share
The table gives the entry point or usage basis, not a common unit. As read on 17 November 2026: Supermemory's Scale, $399 a month, has "Unlimited team seats, per-tag access"; Basic Memory allows "up to 2 people on Cloud, 5 on Pro, and 25 on Business", at $19, $99 and $249 a month per workspace; Cognee charges $5 a month per additional workspace; Hindsight has "no fixed monthly fee, no seat pricing". As read on 18 November, Mnemoverse rooms "carry no separate price": usage "is always billed to the authenticated caller, never the room owner", with any API key, Free included.
This page's reading: what a plan meters differs from vendor to vendor: seats, workspaces, projects, MCP seats or tokens. Some entry points carry no seat price at all: at Cognee "The first Personal Workspace is free" and "workspaces support unlimited users", and Cloudflare says "We are not currently billing for Agent Memory during private beta."
How shared memory leaves: export and moving between instances
This section covers leaving an instance of the same product, not moving between vendors.
Hindsight's post of 22 September 2026 describes moving a whole bank: "Export and import a whole bank as a ZIP archive with an explicit scope: data, bank config, history." It calls this "the cross-instance path" and says "The endpoints are live on Hindsight Cloud"; Vectorize's pricing page adds: "Contact us for assistance migrating your memory data from a self-hosted instance to Hindsight Cloud." Caveats: merge mode "doesn't dedupe facts" and a restore target "must not already exist"; Hindsight's memory-banks API page adds: "Aliases are not carried by export, import or clone."
Supermemory: "Organization exports require owner or admin access, and download links expire after six hours." The table gives the routes for Mem0, Basic Memory, Cognee, Letta and Claude, and AWS's record operations; the self-hosted Company Brain's "Export as Markdown" is in the Supermemory section above. For an account's own memory, Mnemoverse exports every memory as JSONL, Markdown, CSV or text, and its JSONL file "is also what Bring your memory reads back".
On the pages read, Cloudflare's statements are a promise and a future plan, not a described current export workflow: "Every memory is exportable", and "In the future, snapshots and exports will go to R2 for cost-efficient long-term storage."
Where it can run
Among the options in the table, "Hindsight Cloud is shared and multi-tenant", Basic Memory self-hosts under AGPL-3.0 and Caura's engine is Apache 2.0, while Supermemory Local "is not open source".
Letta's shared memory repositories work with agents hosted in Letta Cloud. Hindsight's Enterprise features "can be turned on for a Cloud organization", so SSO does not imply a dedicated deployment. Self-hosted Claude Managed Agents is available on request as of 17 November 2026, unlike self-hosted sandboxes, where "you host only the sandboxes, and Anthropic runs the rest." For residency, see "Where may the data live?" in the cross-IDE comparison.
Where ChatGPT shared projects fit
A ChatGPT shared project shares a project's chats, files and instructions with invited members, under project-only memory. OpenAI's help article, read 17 November 2026, puts project sharing on Free, Go, Plus, Pro, Business, Enterprise and Edu: "ChatGPT can draw from anything in the shared project", and "Invitees receive either chat or edit access." A person's saved memories are separate: the Memory FAQ for Business, read 13 November 2026, calls them "Not transferable to other users, even within the same Business workspace."
On the community forum, Business user Blake Thomas wrote on 2 August 2026: "Memory changes from Default to Project-only." On 8 September OpenAI Support, marked as staff, answered around a link to the help article: "Shared Projects already share project chats, files and instructions, but they remain project-only, and Work isn't available in that mode" and "We'll pass along the request for governed shared Work and Library access; no timeline to share."
This page's reading: ChatGPT shares memory at the level of a project inside the assistant; it is not a memory layer a developer connects agents to.
What to check before choosing a shared memory layer
The table answers what each vendor documents. Before a team commits, four questions are worth putting to the vendor's own pages:
- Is the shared container an authorization boundary at retrieval, or a billing and organizing object?
- When a member leaves, are the API keys and agent identities they created revoked too?
- What does an export carry: memories only, or also configuration and history?
- Does the self-hosted edition include the shared-space feature itself?
What this page does not claim
No product was tested and no account was opened on any plan. The five questions are this page's framing; lines marked "This page's reading" are interpretations, not vendor claims. Behaviours, plans and prices are the vendors' documented ones on the date in Sources. Statuses on 17 November 2026: Foundry memory, public preview; Cloudflare Agent Memory, private beta; Managed Deep Agents, public beta; Claude Managed Agents and memory stores, beta, self-hosted on request; Zep SCIM, beta. On 18 November, Mnemoverse's console export and import pages are in beta. Where a vendor's pages say nothing about a question, the table says so with the pages read; that is not a claim the product does without it. Supermemory's Company Brain is described only from Supermemory's own blog, changelog and repository. The Mnemoverse row is taken from Mnemoverse's own documentation, read 18 November 2026, by the same rule as every other row.
Common questions
List AI agent memory layers with support for shared agent workspaces or multi-user rooms.
These systems document a shared space for a team, as read on 17 November 2026: Supermemory (organization, spaces, namespaces), Mem0 (organizations, projects), Zep (account, projects), Letta (organization, shared memory repositories), Cognee (workspace as tenant), Hindsight (organization, memory banks), LangSmith with Deep Agents (organization memory as a namespace you define; agent memory shared across a Managed Deep Agents deployment), Claude Managed Agents (workspace memory stores, in beta), Google Agent Platform Memory Bank (scopes), AWS AgentCore Memory (namespaces), Microsoft Foundry memory (scopes, in public preview), Basic Memory (team workspaces), Sentra (organization memory), Caura (agent, fleet and tenant scopes) and Cloudflare Agent Memory (shared profile, in private beta). Mnemoverse, this page's publisher, shares memory through rooms across accounts (its own documentation, read 18 November 2026).
What did Supermemory discontinue on 9 September 2026, and what still runs?
On 9 September 2026 Supermemory stopped its hosted Company Brain and Nova and refunded everyone who was charged; its MCP and plugins keep running. On 25 September it open-sourced the Company Brain harness for your own Cloudflare account. That build's guide says it "doesn't expose an MCP server or an API for asking the brain questions", and the build can export memories as Markdown.
Which agent memory layers let a team manage members with SSO or SCIM?
SSO is often an Enterprise feature: Hindsight, Letta, Supermemory and Sentra list it there, LangSmith puts SAML SSO and SCIM on Enterprise, and Basic Memory sells SSO as a Business add-on. At Zep, "SCIM provisioning is in beta. It requires the Enterprise plan. Zep turns on SCIM for each Enterprise account on request." For Zep's Memory MCP Server: "Custom OIDC and UserGroup ABAC require the Enterprise plan. Google Workspace sign-in does not." All as read on 17 November 2026. Mnemoverse, as read on 18 November, lists SSO/SAML and SCIM on its Enterprise plan.
Does ChatGPT share memory across a team?
Yes, within a shared project. OpenAI says "ChatGPT can draw from anything in the shared project" and "Shared projects automatically use project-only memory and cannot access an individual member's memories or context from outside the project" (17 November 2026). A person's saved memories are separate: OpenAI's Memory FAQ for Business, read 13 November 2026, calls them "Not transferable to other users, even within the same Business workspace."
Can a team take its shared agent memory with it?
It depends on the vendor. Hindsight exports "a whole bank as a ZIP archive" and calls this "the cross-instance path"; Supermemory exports JSON "of up to 25,000 documents and memories"; Mem0's exported data "expires after 7 days"; Basic Memory exports Markdown and project ZIPs; Cloudflare's blog promises "Every memory is exportable" for a product in private beta, and its R2 snapshots and exports are a future plan. All as read on 17 November 2026.
Sources
Grouped by vendor. Every page was read on 17 November 2026, except OpenAI's Memory FAQ for Business, read on 13 November 2026, and Mnemoverse's own pages, read on 18 November 2026.
- Supermemory, read 17 November 2026: An update to supermemory (blog, 10 September 2026); Open-sourcing Company Brain (blog, 25 September 2026); changelog entries Introducing Company Brain (24 July 2026), Company Brain on Max (13 August 2026), Company Brain is open source (25 September 2026) and data export from Settings (4 June 2026); changelog; pricing; security; invite team members; MCP and spaces; multi-tenancy examples; Local and Enterprise; billing; security overview; connectors; documentation export (153 pages); Company Brain repository; outside-Slack guide.
- Mem0, read 17 November 2026: pricing; platform pricing; organizations and projects; add organization member; Platform and open source; exporting memories; open source to Platform; changelog highlights; documentation index.
- Zep, read 17 November 2026: Agent Memory MCP (blog, 19 August 2026); pricing; role-based access control; Enterprise SSO; SCIM provisioning; UserGroup access; agent access; governance; FAQ;
zepctlCLI; Graphiti graph namespacing. - Letta, read 17 November 2026: pricing; collaborators; permissions; teams quickstart; shared memory; CLI reference; self-hosting.
- Cognee, read 17 November 2026: pricing (SSO by engagement read from the comparison table's cell labels); data isolation and multi-tenancy; permissions and access control; members; dataset Markdown export, API and Cloud API.
- Hindsight (Vectorize), read 17 November 2026: Move a memory bank (blog, 22 September 2026); memory banks; admin CLI; Hindsight Cloud; team management; Cloud; Enterprise; SSO; bank transfer API; Vectorize pricing.
- LangChain (LangSmith and Deep Agents), read 17 November 2026: pricing; RBAC; user management; organization hierarchy; authentication methods; Deep Agents memory; Managed Deep Agents memory.
- Anthropic (Claude Managed Agents), read 17 November 2026: memory; memory stores API; workspaces; user management; Managed Agents overview; self-hosted quickstart; self-hosted data and access; pricing.
- Google (Agent Platform Memory Bank), read 17 November 2026: Memory Bank; generate memories; fetch memories; pricing.
- AWS (Bedrock AgentCore Memory), read 17 November 2026: long-term memory organization; agent authorization; disaster recovery; cross-account access; pricing.
- Microsoft (Foundry Agent Service memory, preview), read 17 November 2026: memory usage; what is memory.
- Basic Memory, read 17 November 2026: pricing; security; about teams; members and roles; team projects; seats and billing.
- Sentra, read 17 November 2026: llms.txt; pricing.
- Caura, read 17 November 2026: agent memory; repository.
- Cloudflare (Agent Memory, private beta), read 17 November 2026: Introducing Agent Memory (blog, 17 April 2026, modified 15 July 2026); documentation; pricing; namespaces and profiles.
- OpenAI (ChatGPT shared projects), read 17 November 2026: Projects in ChatGPT; Managing data, sharing, and privacy in ChatGPT Business; More ways to work with your team (25 September 2025); community forum thread (2 August to 8 September 2026). Memory FAQ (Business Version): read 13 November 2026.
- Mnemoverse (publisher), read 18 November 2026: rooms; pricing; export; import.
- Practitioners, read 17 November 2026: snowhy, Ask HN (9 July 2026); Naman Jain, dev.to (12 November 2026).
Related
- Cross-IDE context platforms for teams
- State of AI Agent Memory 2026
- Shared memory for multi-agent systems
- MCP servers sharing state across IDEs
- Rooms: shared memory across accounts
Sources were read on 17 November 2026, OpenAI's Memory FAQ for Business on 13 November, and Mnemoverse's own pages on 18 November; the page describes them as of those dates.
Who wrote this. Mnemoverse publishes this page and is one of the systems it compares, so we are an interested party. That is why every vendor fact here comes from a page listed under Sources with the date it was read, and why the Mnemoverse row is taken from Mnemoverse's public pages in the same way as every other row. If you find a fact that is wrong or out of date, tell us and we will correct it.
Edward Izgorodin · Mnemoverse · 18 November 2026
