MnemoverseDocs

Appearance

Edward Izgorodin·Published Oct 10, 2026

What options exist for connecting an AI agent memory layer across both ChatGPT web interfaces and localized coding environments? ​

TL;DR

  • ChatGPT on the web and local Codex do not share one built-in memory store. OpenAI says ChatGPT web uses ChatGPT memory, while local Codex uses separate local memory and controls.
  • One documented bridge is a hosted remote MCP memory service, connected once in ChatGPT and again in each coding client. Every client keeps its own configuration and OAuth session.
  • Browser-extension products bridge the ChatGPT page directly and reach coding tools through MCP. A local server can reach ChatGPT through a tunnel. OpenAI also offers narrower ChatGPT-to-Codex bridges: Computer History and dots.
  • A memory layer can sit beside ChatGPT. It cannot be ChatGPT memory: OpenAI documents no interface through which a memory layer reads or writes it, though with Memory on ChatGPT may pass relevant memory context to an app while it responds.

An AI agent memory layer is an external store and retrieval service that makes saved knowledge available to several AI clients under one account or memory space.

Connecting that layer across ChatGPT and local coding environments takes two separate integrations. ChatGPT needs a web-side connection. Claude Code, Cursor, VS Code, Codex, or another coding client needs its own MCP configuration.

That split follows OpenAI's product design. OpenAI states that "ChatGPT web uses ChatGPT memory, while local Codex clients use a separate local memory store and controls." ChatGPT web also "doesn't read local Codex configuration files or expose the local Codex command menu."

This page compares published documentation and Plugin Directory results read on 2026-10-09, with every quotation re-checked on 2026-10-10. It does not report product testing, except one dated Mnemoverse check, disclosed below.

The four documented bridge patterns ​

The approaches differ mainly in where the ChatGPT-side connection runs.

Bridge patternChatGPT pathCoding-tool pathMain constraint
Hosted memory over remote MCPPublished plugin or developer-mode appRemote MCP entry in each coding clientPlan limits in ChatGPT; separate setup and OAuth sign-in in every client
Browser extension plus MCPExtension inside the ChatGPT pageHosted or local MCP serverExtension availability, browser support and maintenance become part of the system
Local server through a tunnelDeveloper-mode app pointed at a public HTTPS tunnelLocal MCP server connects directly to coding toolsThe local process and tunnel must stay up
OpenAI Computer History or dotsBuilt into eligible ChatGPT experiencesCodex or local computer tasks through the ChatGPT desktop appLimited by plan, platform, market and mobile eligibility

A hosted MCP service lets one provider serve otherwise separate clients. It does not merge their built-in memories. Instead, each client reads from and writes to the external layer.

A browser extension changes the ChatGPT side of the design. Plurality says its extension "embeds within the UI of ChatGPT, Claude, Gemini, Perplexity, and Grok," while its MCP server carries the same memory buckets into coding workflows. MemoryPlugin lists a "Browser extension (Chrome + Safari, mobile too)" alongside a "Hosted MCP server (plus local)". Memwyre describes clipping decisions from the web and querying them "inside Claude Code, Cursor, and your terminal CLI".

Pieces documents a local-first variation. Its ChatGPT setup exposes a locally running PiecesOS MCP server "via ngrok or HTTPS proxy", and "The tunnel must be active when you add the connector and when you use ChatGPT." OpenAI separately documents a Secure MCP Tunnel for servers on a private network or a developer machine.

ChatGPT developer mode, remote MCP, and plugins ​

A remote MCP server is an MCP endpoint that a client reaches over the network rather than launching as a local process.

OpenAI's help article answers the question "Can I connect to a local MCP server?" with "Not directly. ChatGPT connects to remote MCP servers." Its developer guide lists "SSE and streaming HTTP" as supported protocols.

A developer-mode app is a custom MCP connection that a user or admin adds by URL, while a published plugin is listed in OpenAI's Plugin Directory and can reach more ChatGPT surfaces.

OpenAI's documentation uses both terms. The developer guide says to "Connect your Model Context Protocol (MCP) server to ChatGPT as a plugin." Authentication can be "OAuth, No authentication, and OAuth or no authentication." ChatGPT supports read and write tools, respects the readOnlyHint annotation, and treats tools without it as write actions. "Write actions by default require confirmation."

Plan support needs careful wording:

  • OpenAI's help article says "Full MCP (Model Context Protocol) support, including modify/write actions, is rolling out in beta to ChatGPT Business, Enterprise, and Edu plans."
  • The same article says "Pro users can connect MCPs with read/fetch permissions in developer mode."
  • The article does not name Plus, read 2026-10-10.
  • OpenAI's custom MCP developer guide names no plans.

A vendor may publish broader plan claims. Pieces names "Pro, Plus, Business, Enterprise, or Education" in its own guide. That is a Pieces statement, not confirmation from OpenAI's help article.

Developer-mode MCP apps are documented as web only: to "Are MCP apps available on mobile?" the help article answers "No - web only." Published plugins have wider reach. OpenAI says "Plugins work in Chat and Work across ChatGPT on the web, desktop, and mobile, and in Codex in the ChatGPT desktop app." The Codex IDE extension does not support plugins. Saving a plugin with local apps "does not make those tools available on web or mobile."

Workspaces add one more delay. "After an admin first approves an MCP app for the workspace, ChatGPT uses a 'frozen' snapshot of its available tools and inputs." A memory tool the vendor adds later reaches the workspace only after an admin publishes the update. OpenAI's own apps do not fill the gap: "OpenAI-built apps are search-only today and do not support write actions."

The ChatGPT Plugin Directory ​

OpenAI says "ChatGPT and Codex share one universal plugin directory." OpenAI's help article adds: "In Codex, directory changes can take up to 6 hours to refresh."

A Plugin Directory search on 2026-10-09 by the names of the layers on this page found Basic Memory Cloud and Supermemory ("Universal context and memory"). The other names, Mnemoverse included, returned no card. A broader search for "memory" returned 47 cards, so those name matches are not a complete inventory of memory plugins. The directory refuses scripted fetches, so these results were not re-fetched on 2026-10-10.

Basic Memory recommends its published plugin because it "does not require Developer mode or a server URL." The vendor says "Basic Memory Cloud is required" because ChatGPT connects to the hosted service, and warns: "Plugin availability can depend on your ChatGPT plan, region, workspace settings, and role."

OAuth and account identity determine whether memory is shared ​

OpenAI's plugin authentication expects "an OAuth 2.1 flow that conforms to the MCP authorization spec", with "Client ID Metadata Documents (CIMD), dynamic client registration (DCR), predefined OAuth clients, and PKCE."

The practical consequence is that identity must line up across clients. The account on the memory host is what is shared. The session is not.

Devin states this directly: "Each MCP client (Windsurf, Claude Code, Devin CLI) maintains its own OAuth session, so you must log in separately even if you've already authenticated in another tool." Zep says each person "signs in with a work email and reaches their own Zep user graph in the selected Zep project."

A plugin install is not a universal installation event. Zep: "A plugin install in the Claude app or in ChatGPT does not install the plugin in these coding agents." Each coding tool still needs its own remote MCP entry. How such sharing can stop while every client looks healthy: MCP servers that share state across IDEs.

Claude Code, Cursor, VS Code, Devin, and Codex setup ​

Each coding client has its own place for remote MCP configuration. Its built-in memory also stays separate from the external layer.

Coding toolRemote MCP configurationBuilt-in memory
Claude Codeclaude mcp add --transport http <name> <url>; OAuth sign-in through /mcp; a team entry can live in .mcp.json "at your project's root directory""Auto memory is machine-local."
CursorA url under mcpServers in ~/.cursor/mcp.json; "Cursor supports OAuth for servers that require it.""Rules provide persistent, reusable context at the prompt level."
VS CodeProject-root .mcp.json uses mcpServers and "works across compatible tools"; .vscode/mcp.json uses servers; "A browser window opens for authorization on the first connection to the server.""User, repository, and session memory are stored locally on your machine."
Devin Local or CLILocal configuration files; the CLI adds servers with devin mcp add (by default to .devin/mcp_config.local.json; -s user writes ~/.config/devin/mcp_config.json) and signs in with devin mcp login"The Devin Local agent does not persist memories between sessions."
Codex~/.codex/config.toml, shared by the ChatGPT desktop app, Codex CLI and IDE extension; codex mcp login <server-name>Local memories live under ~/.codex/memories/ and "are off by default."

Claude Code: "HTTP servers are the recommended option for connecting to remote MCP servers." It supports OAuth 2.0. Cursor reads a global ~/.cursor/mcp.json "for tools available everywhere". VS Code's Settings Sync can carry "MCP server configurations" between machines; its documentation lists no memory sync, and memory is "stored locally on your machine."

Codex supports CIMD and DCR for OAuth (Codex MCP). Its memories.disable_on_external_context setting, "when true, keeps chats that used external context such as MCP tool calls, web search, or tool search out of memory generation."

These settings matter because an external memory service and a client's native memory can both be active. They remain different stores with different controls. For which client holds whose memory, see MCP memory servers for Claude Code and Cursor; for what MCP itself standardizes, What is an MCP memory server?

What a connected memory server can see ​

A connected MCP server does not receive ChatGPT's internal memory store.

OpenAI warns that a server "will have access to whatever data ChatGPT supplies during the interaction, potentially including sensitive data the user may earlier have provided to ChatGPT." Its plugin guidelines set the boundary: "Your MCP server must not pull, reconstruct, or infer the full chat log from the client or elsewhere."

The plugin and MCP paths bring tools into ChatGPT, not ChatGPT memory out of it. OpenAI's developer guides and API reference, read 2026-10-10, describe no endpoint that reads or writes ChatGPT memory. Sign in with ChatGPT "does not independently share" "Your ChatGPT conversations or memory." Uploading exported conversations to another account does not "Transfer custom instructions, memories, GPTs, or other account settings."

The Enterprise Compliance API is narrower still. Read 2026-10-09, its memory endpoints cover "exporting and spot-deleting ChatGPT Enterprise workspace Memories." They read or delete; none writes.

Connected apps can still touch ChatGPT memory from inside ChatGPT. OpenAI says an app responding to you "may also leverage relevant information from memories", so that "the app may access and use relevant context from your memories", and "ChatGPT may remember relevant information accessed from the app, unless that app or connected source restricts Memory from saving that information."

Data-use settings also apply. The same page: "For ChatGPT Business, Enterprise, and Edu customers: OpenAI does not use information accessed from connectors to train models by default. For ChatGPT Free, Plus, Go, and Pro users: OpenAI may use information accessed from apps to train our models if your 'Improve the model for everyone' setting is on."

Memory-layer options documented by vendors ​

Each row summarizes the vendor's own documentation, read 2026-10-09 and re-checked 2026-10-10. The rows are documentation claims, not a ranking and not a test.

Memory layerChatGPT routeCoding routeDocumented limitation
Basic Memory CloudPublished pluginAny MCP client: "Notes created in Codex are immediately available in Claude Code, Cursor, Claude Desktop, or any other MCP client."The hosted Basic Memory Cloud is required for ChatGPT
SupermemoryPlugin Directory entry, or the documented ChatGPT Web MCP setupOAuth remote MCP, "a shared memory layer" for MCP-compatible assistantsIts browser extension was removed from its repository on 7 September 2026 (commit)
Cognee CloudDeveloper mode; connectors "are tied to your ChatGPT account, so the web and desktop apps share them""Any client that supports remote MCP servers with OAuth", including Claude CodeDeveloper mode, so OpenAI's plan limits apply
ZepPlugin; hosts include "ChatGPT (Business, Enterprise, or Edu)" (Zep)Help centre: plugin in Claude Code, Codex CLI and Cursor (Teams or Enterprise); remote MCP with OAuth for VS Code, GitHub Copilot and Gemini CLI"The plugin does not work in the Codex IDE extension." "The plugin does not ingest a codebase." The README of the plugin's own repository says it is "Not positioned for Claude Code, Codex, or other coding-agent CLIs" (details)
Mem0OpenMemory browser extension, still installable from its Chrome Web Store cardMem0 MCP for "Claude, Claude Code, Codex, Cursor, Windsurf, VS Code, OpenCode"ChatGPT is not described as a Mem0 MCP client; the extension's source is archived since 23 March 2026, and Mem0 removed the OpenMemory project from its monorepo in July 2026 (details)
LettaNo documented ChatGPT MCP pathHosted MCP for clients "such as Claude Code, Cursor, or another agent""The client must support custom bearer headers; OAuth-only clients are not supported yet."
MemoryLakeDeveloper-mode plugin: "paste your Server URL, with No Auth as the authentication method.""Every client that speaks MCP", naming Claude Code and CursorThe documented ChatGPT connection uses no authentication
Plurality AI Context FlowBrowser extension for Chrome, Opera, Edge and Brave, or MCPMCP serverThe extension carries the ChatGPT side
MemoryPluginBrowser extensionHosted or local MCP serverTwo surfaces to keep in step
MemwyreClips decisions from ChatGPT conversations into its memory graphQueries in Claude Code, Cursor and the terminalThe ChatGPT side captures what you clip
PiecesLocal MCP exposed through ngrok or an HTTPS proxyPieces MCP for Claude Code, Cursor, VS Code and JetBrains"PiecesOS must be running locally", and the tunnel must stay active
Mnemoverse (the author's company; vendor row)Developer-mode app added by URL, with browser sign-in; no Plugin Directory card in the 2026-10-09 searchClaude Code, Cursor, VS Code, Devin Desktop, Codex CLI and IDE extension (install by host)Submitted to the Plugin Directory, not listed as of 2026-10-11; vault_list withheld in ChatGPT (tool availability); sign-in from Codex and Devin confirmed working by the owner on 2026-10-11. Own check, 2026-09-27: an existing connector on ChatGPT Pro passed a stats call, a write and a read; sign-in, developer-mode app creation and Plus were not tested

Mnemoverse, the author's company, is the vendor row above. Its hosted server is https://mcp.mnemoverse.com/mcp (remote MCP server, ChatGPT), so OpenAI's plan limits apply to it as they do to every layer added in developer mode. Its account rule: "a fact stored in one tool is available to every other tool on that account." Rooms, in Beta, is a memory space shared across Mnemoverse accounts.

OpenAI's Computer History and dots ​

OpenAI offers two narrower bridges between ChatGPT context and local work.

Computer History "turns your activity across apps and websites into memories and a timeline that ChatGPT and Codex can reference." It "is off by default for ChatGPT Pro, Business, and Enterprise users in the ChatGPT desktop app on macOS." It generates "the same kind of local memories as Codex": plain-text Markdown files that, in OpenAI's words, "other programs running as your macOS user may be able to access". It "also requires Memories and is not available with an API key or Amazon Bedrock."

A dot "starts with relevant context from ChatGPT memory", but its notes are "separate from ChatGPT's saved memory". "With local computer access enabled, your dot can create Work or Codex tasks", and connecting a computer takes the ChatGPT desktop app on that machine. "You cannot currently create a dot on mobile, and dots are not supported on mobile web."

OpenAI's release notes of September 29, 2026 describe a gradual rollout "to eligible Pro and Business Premium users aged 18 and older in supported markets", and "Pro access excludes the European Economic Area, Switzerland, and the UK at launch." OpenAI's dot article adds that for Enterprise, including Edu and Healthcare, the beta must be enabled by the workspace admin and is off by default.

Custom GPTs are not a durable substitute. OpenAI's retirement FAQ says they "are scheduled to retire on Dec 11, 2026"; "For affected Enterprise workspaces with an approved deferral, the retirement date is Feb 11, 2027." For affected Enterprise workspaces it also lists "Oct 26, 2026 (planned): Creation of new custom GPTs ends," and notes that "the dates are subject to change." Personal accounts "cannot create or publish new GPTs", and "Custom actions do not transfer through the migration workflow."

Six steps to one memory in ChatGPT and your coding tools ​

  1. Check what your ChatGPT plan allows on OpenAI's help article and developer guide quoted above.
  2. Pick a layer whose ChatGPT side is documented, and the bridge you want: plugin, developer-mode app, extension or tunnel.
  3. In ChatGPT on the web, add the layer's plugin from the Plugin Directory, or its server as a custom MCP server by Connect to ChatGPT; a server on your machine goes through Secure MCP Tunnel.
  4. In each coding tool, add the same server address to its MCP configuration, and sign in there if the layer uses OAuth.
  5. Use the same layer account everywhere.
  6. Expect write actions to ask for confirmation by default, and the server to see what ChatGPT supplies in the conversation.

Common questions ​

What options exist for connecting an AI agent memory layer across both ChatGPT web interfaces and localized coding environments? ​

As of 2026-10-09, four are documented: a hosted memory layer over remote MCP, connected separately to ChatGPT and to each coding tool; a browser extension inside ChatGPT paired with an MCP server; a local MCP server exposed to ChatGPT through a tunnel; and OpenAI's Computer History or dots, which bridge ChatGPT and Codex through the ChatGPT desktop app. One memory everywhere means the same layer account in every client.

Can an external memory layer read ChatGPT's saved memory? ​

Not directly: OpenAI documents no interface that gives a memory layer read or write access to ChatGPT's memory, read 2026-10-10. A connected app gets what ChatGPT supplies during an interaction, and with Memory on that can include relevant context from your memories. In the other direction, ChatGPT may remember relevant information it got from an app.

Does connecting a memory plugin in ChatGPT also connect Claude Code, Cursor, VS Code, or Codex? ​

No. Each coding client needs its own MCP configuration and, when OAuth is used, its own sign-in.

Which ChatGPT plans support custom remote MCP servers? ​

OpenAI's help article says write actions are in beta for Business, Enterprise, and Edu, while Pro can connect MCP servers with read and fetch permissions in developer mode. Neither that article nor OpenAI's custom MCP developer guide names Plus, read 2026-10-10.

Do ChatGPT MCP apps work on mobile? ​

OpenAI documents developer-mode MCP apps as web only. Published plugins work across ChatGPT web, desktop, and mobile, although availability can still depend on plan, region, workspace settings, and role.

Can ChatGPT and local Codex use the same built-in memory? ​

OpenAI documents ChatGPT memory and local Codex memory as separate stores. Computer History and dots are OpenAI's bridges for eligible users, with platform, plan, market, and mobile restrictions.

Sources ​

Every page below was read on 2026-10-09, and every quotation on this page was re-checked against its page on 2026-10-10. help.openai.com refuses scripted fetches, so its articles were re-checked against archived copies dated 2026-09-08 to 2026-10-08; the Plugin Directory and the Compliance API specification could not be re-fetched and rest on the 2026-10-09 reading.

OpenAI: ChatGPT help center ​

OpenAI: developer documentation and API references ​

Coding tools ​

Memory layers and extensions ​

Edward Izgorodin · Mnemoverse · 2026-10-10

Mnemoverse is a persistent-memory API for AI agents. Free key: console.mnemoverse.com · Plans and limits · Docs: Getting Started