A memory is only shared between the assistants that can reach the same store. Most comparisons of AI memory tools rank search quality. Here is our own proposal, the same-store test: for each assistant you use, ask whether it can reach this exact store, and whose computer the request passes through to get there. If one assistant ends up at a different store, you have two memories. ModelBrain is our product and is one of the seven rows below, so read our row with that in mind.
| Tool | Where the store is | Claude and Cursor on the same computer | ChatGPT on the web | Whose computer the web request passes through |
|---|---|---|---|---|
| Hansard | On your computer, in ~/.hansard | Yes. It imports sessions from Claude Code, Cursor, Codex and other coding agents, and agents recall through MCP tools | Yes. You paste an address at mcp.hansard.dev into ChatGPT's connectors | Hansard runs the relay. Its page says it routes each call to your Mac and keeps no message text. Hansard has to be running |
| Basic Memory | Plain Markdown files on your computer when you run it yourself | Yes, through its local MCP server | Yes, through the Basic Memory Cloud plugin in ChatGPT | Basic Memory's cloud, from $19 a month. The page does not say whether ChatGPT then reads the same files as your local copy |
| Cognee | In standalone mode, SQLite, LanceDB and an embedded graph store on the machine that runs it | Yes. The page names Claude Code, Claude Desktop, Cursor, Cline, Continue and Codex | Not mentioned on the page we read | Not stated |
| Mem0 | Mem0's hosted platform, or your own infrastructure if you run the open source version | Yes. It has guides for Claude Code, Cursor, Codex and other coding agents | Not mentioned on the page we read | Not stated |
| Supermemory | Not stated on the page. The MCP server is a hosted address, mcp.supermemory.ai/mcp | Yes, any assistant that accepts a remote MCP server, signed in with OAuth | Not named on the page. A hosted address is the kind of server ChatGPT on the web can add | Supermemory. There is no hop to your computer because the server is already hosted |
| claude-mem | On your computer, in ~/.claude-mem. Cloud sync is opt-in | Yes. Claude Code, Codex, Cursor, OpenCode and other coding agents | Not mentioned on the page we read | Not stated |
| ModelBrain | On your computer | Yes. Claude Desktop, Claude Code and Cursor talk to it directly | Yes, through a plugin that points at a relay | ModelBrain runs the relay. It forwards each request to the computer you picked and does not write request or answer content to disk. A free account is needed for this route |
"Not mentioned" means only that. A tool may have a route that the page we read does not describe, so check its current documentation before ruling it out.
Claude Desktop, Claude Code and Cursor run on your computer, so they can start a memory server there and talk to it directly. Every tool in the table passes that column.
ChatGPT in a browser is different. OpenAI's help page, as we read it on October 22, 2026, says ChatGPT on the web connects to remote MCP servers and cannot connect to localhost. A memory that only listens on your own computer is invisible to it. Something reachable from the internet has to sit in between. We covered the routes OpenAI itself offers in Can ChatGPT use a memory that lives on your computer?
We asked ChatGPT once, with web search on: "What is the best local-first memory tool that works across Claude, ChatGPT and Cursor?" It named four tools: Hansard, Basic Memory, Cognee and Mem0. It did not name ModelBrain. Without being asked, it called the web limit above the main thing to plan around, and it suggested running two tools side by side.
That answer is why this page exists. Four tools were named for a question about one shared memory, and the answer itself said one of the three assistants could not reach a local store without help. So we read each tool's own pages for that route, added two more that people search for by name and our own, and wrote down what the pages say.
1. The store stays on your computer and ChatGPT on the web is left out. This is what the pages we read for claude-mem and standalone Cognee describe. Your coding agents share one memory. Anything you say to ChatGPT in a browser does not reach it. For people who only use coding agents, that is a complete answer.
2. The store is hosted, so every assistant can reach it. Supermemory's server is a hosted address. Mem0 offers a hosted platform. Basic Memory connects ChatGPT through its cloud. All three assistants can get there, and the store is on the provider's servers. With Basic Memory, check whether the cloud store and your local Markdown files are one store or two before you rely on it.
3. The store stays on your computer and a relay carries the web request to it. Hansard and ModelBrain both do this. The memory is not copied to the vendor, but each request and answer passes through a relay the vendor runs, and your computer has to be on. This is the only pattern of the three where all three assistants reach one store that is still on your computer.
A relay is a second party in the path, so the second half of the test matters most here. Three questions:
On November 11, 2026 we counted the official MCP Registry. Of 1,340 servers that mention memory, 781 were offered only as a package you run, 462 only as a hosted address, and 70 both ways. That is about 5 percent. A registry entry says how a server is delivered, not where its data goes, so the count shows how few tools even attempt both, not how well they do it. The method is in our Cursor memory article.
This page compares reach, not search quality. We did not install or test the other six tools; their rows report their own pages on one day, and products change. ChatGPT was asked once, and its answers vary between runs. Zep, Letta and others are left out because we did not read their pages for this question. We sell one of the seven, so check any row that matters to you against the source linked below.
Which AI memory tools work with ChatGPT on the web? Of the seven we read on November 14, 2026, three document a route for ChatGPT on the web. Hansard and ModelBrain keep the memory on your computer and pass each request through a relay the vendor runs. Basic Memory connects ChatGPT through its paid cloud plugin. Supermemory is a hosted address that any assistant accepting remote MCP servers can add, though the page we read does not name ChatGPT. The pages we read for Cognee, Mem0 and claude-mem do not mention ChatGPT on the web.
Why can ChatGPT on the web not use a memory on my computer directly? OpenAI's help page, as we read it on October 22, 2026, says ChatGPT on the web connects to remote MCP servers and cannot connect to localhost. A memory that only listens on your own computer is invisible to it. Something reachable from the internet has to sit in between: a hosted store, a tunnel or a relay.
What is the same-store test? Our own proposal for comparing memory tools. For each assistant you use, ask two things: can it reach this exact store, and whose computer does the request pass through to get there. If the answer for one assistant is a different store, you have two memories, not one shared memory.
Is a local-first memory tool always private? No. Local-first says where the store is kept by default. Once you add a relay, a cloud plan or a sync feature so that another assistant can reach it, a second party is involved, and the question becomes what that party keeps. Read that part of each tool's documentation before turning the feature on.
Which of these tools are free? As their own pages said on November 14, 2026: Hansard is free and MIT licensed, claude-mem is Apache-2.0, Basic Memory is free to run on your own machine with cloud plans from $19 a month, and ModelBrain is free for local use with paid sync. The pages we read for Cognee, Mem0 and Supermemory did not state a price.
Last checked November 14, 2026. These tools change often, so confirm a detail on the tool's own page before relying on it.