All tools / Guides

How to share one memory across your AI bots

ChatGPT, Claude and Cursor each keep their own memory, and none of them read the others'. The common fix is one memory server that every bot connects to over the Model Context Protocol (MCP). Here's how to set that up without losing data. Last reviewed 2026-11-09.

1. Put the memory in one place

Pick a single memory server (hosted or self-hosted) and connect every client to it. Don't give each app its own copy. MemoryStack tracks 66 memory tools. Start with the ones whose docs name all the apps you use: 5 tools name ChatGPT, a Claude app and Cursor (see the list).

2. Check how each app connects

Desktop and coding apps (Claude Desktop, Claude Code, Cursor, Codex, VS Code) can usually launch a local MCP server as a process (the stdio transport) or connect to a remote one.
ChatGPT on the web adds a custom MCP server by entering a server URL (or a Secure MCP Tunnel). It supports the SSE and streaming HTTP protocols, with OAuth or no authentication (OpenAI docs). A stdio-only server on your laptop isn't enough. To include ChatGPT, your memory needs a remote endpoint it can reach.

On every tool page, the MCP transport row shows which of these a tool documents.

3. Never let two processes write the same memory file

The official reference memory server stores everything in one JSONL file. If two clients each start their own copy pointed at the same MEMORY_FILE_PATH, they can silently discard each other's writes, because the lock only works inside one process (issue #4797). Safer setups:

Run one server process and point every client at it over HTTP.
Use a server backed by a real database (SQLite, Postgres, Neo4j and so on). Each tool page lists the storage it uses.
Or keep separate memories per client on purpose. Cognee MCP, for example, gives each client its own dataset by default (Cognee MCP README).

4. Set the storage path explicitly, then check it

Users of the official memory server reported that a custom MEMORY_FILE_PATH was ignored and memory.json ended up in npx's temporary directory, where it can vanish (issue #692). After setup, save one memory and confirm the file or database is where you expect.

5. Importing is not syncing

Claude can import memories from another AI service: you paste an export into Claude's import flow and it extracts entries. That's a one-time copy, not a live link, and Anthropic calls it experimental (Claude Help Center). For memory that stays in sync, use a shared server as above.

6. Pick a tool

Client support on MemoryStack comes from each vendor's own docs. We don't test tools.