Graph-MIND

Your AI forgets everything when the session ends. This one remembers, and it never calls an LLM to do it.

88.8% on LongMemEval. Zero model tokens spent saving memory. Runs on your own PC.

Decide something in Claude Code on Monday. Ask Codex on Wednesday. It knows.

_{A real run, not a mock-up: python demo/record_demo.py records it from a fresh store.}

The part that sounds wrong

Same 40 questions, same answer model, same prompt, same official judge, pre-registered before anyone looked:

Mem0 runs an LLM over your conversations to decide what to keep. Graph-MIND keeps them word for word and finds the right part when you ask. It is cheaper, and in this test it was also more accurate.

Don't believe it? Good. Every question, answer and judge verdict is in research/runs/, and REPORT.md shows how to rerun it, including the runs that went wrong. Or just try it on your own conversations:

pip install graph-mind-memory
graph-mind-install          # connects Claude Code, Codex, Claude Desktop, Cursor, VS Code, ...

What it is

Graph-MIND records every conversation you have with Claude Code, Codex and Claude Desktop, word for word, in a store on your own PC. When a question depends on the past, the model you are using calls Graph-MIND over MCP and gets back a few thousand tokens of the conversations that answer it.

• Nothing is summarised or rewritten. Saving is a database write. No model call, no tokens.
• Search is hybrid. Your words are embedded on your PC (multilingual MiniLM, no API) and fused with keyword search. Korean and English both work.
• Capture is automatic. A background service reads each app's local transcripts, masks secrets, stores the turns and embeds them as they arrive.
• One memory across your PCs. One sentence to the AI on the first PC, one command on the others (see below).
• No cloud of its own. No account, no telemetry; the memory stays on your machines. What leaves them is what your AI app already sends: the recalled turns go to your model's provider with your question.

Benchmarks

All numbers below come from files in this repository: the pre-registrations, each question's answer, and the judge's verdict on it, under research/runs/. The method is in REPORT.md, including the failures and the corrections.

LongMemEval_S: answer accuracy, 500 questions. The answer model is gpt-5-mini; the judge is the official gpt-4o-2024-08-06.

Head to head (table above): the run was pre-registered with code hashes, and nobody had tuned on those 40 questions. Graph-MIND ran its shipped path with the earlier 20-item packet; Mem0 2.2.1 was the latest open-source release on PyPI. Graph-MIND's lead over both is significant (exact McNemar p = 0.002 and 0.003); with 40 questions, each score has an interval of about ±14 points.

Reading other published numbers. They measure different things, so they do not compare directly with the tables above:

• MemPalace's 96.6% is retrieval recall (R@5): is the right session among the five returned? This table measures whether the final answer is correct.
• Mem0's 94.4% is its managed cloud platform, which includes proprietary components. The open source package tested here is a different system.
• Mastra (94.87%), Emergence (86%), Supermemory (85.2%) and Zep (71.2%) report their own setups and answer models. They were not reproduced here.

As far as we know, everything above 85% on that list runs a model over your conversations when it saves them. Graph-MIND does not.

Install

Requires Python 3.10+ (64-bit). Windows, macOS or Linux. Hosting a brain that other PCs join needs Python 3.12 or older on that one PC (its Postgres helper, pgserver, has no newer build yet); everything else, joining included, works on 3.13 and 3.14 too.

What changed in each release: CHANGELOG.md.

pip install graph-mind-memory
graph-mind-install

If graph-mind-install is "not recognized", pip put it in a Scripts folder that is not on your PATH (common with the Windows Python install manager). python -m install runs the same thing. If it stops with a DLL load error, Windows is missing the Microsoft Visual C++ runtime that the embedding runtime needs: install vc_redist.x64.exe, restart, and run the installer again.

or from source:

git clone https://github.com/goyohan0611-png/graph-mind.git
cd graph-mind
python src/install.py

The installer:

• installs the packages (no PyTorch: the embedding model runs on ONNX Runtime);
• registers the MCP server with every app it finds: Claude Code, Claude Desktop (including the Microsoft Store build), Codex, Cursor, Windsurf, Gemini CLI, VS Code (Copilot), Cline and LM Studio;
• starts the capture service at login (Windows Startup folder, a macOS LaunchAgent, or an XDG autostart entry on Linux);
• downloads the embedding model (about 470 MB, once).

Then restart your AI apps. The server is also listed in the MCP Registry as io.github.goyohan0611-png/graph-mind. Running the installer again is safe. Run it again if you move the folder.

To remove it: graph-mind-uninstall unregisters it from every app, removes the login item and stops the capture service. Your memory stays in ~/.graph-mind unless you add --purge. Then pip uninstall graph-mind-memory removes the program.

What it captures

Before anything is stored, these are masked: API keys, tokens from GitHub, AWS, Google and Slack, private keys, passwords inside URLs, values written after password: / api_key= / token:, and any long machine-random string even from a service no rule knows (by its randomness: 91% of random 20-64 character tokens caught; in 246,750 ordinary chat turns it fired 92 times, nearly all on real ids and tokens). Commit ids, hashes and UUIDs are kept, since you recall those on purpose.

Keep masking on if you use a cloud model: a recalled memory goes to your model's provider with the question, so a stored key would leave your PC the first time it is useful. To have Graph-MIND remember a key, tell it where the key is, not the key: "my OpenAI key is in 1Password, under Dev". If your model runs on your own machines and you want keys remembered as typed, add "mask_secrets": false to ~/.graph-mind/config.json (applies from the next message; already masked text stays masked).

Only what you actually send is captured: the service reads each app's transcript, which is written after you press Enter, so a paste you delete before sending never reaches it. A plain password like hunter2 has no recognizable format and is not masked. To remove something,

just ask your AI: "delete the password I typed earlier". It lists masked candidates (wifi password is ****), deletes only the ones you pick, and the conversation about deleting is not saved. Or, from a terminal:

graph-mind-forget          # asks for the phrase without showing it, then confirms

Either way, what you pick is deleted from this PC and its search indexes, from the shared memory, and from your other PCs on their next sync; only ids are shared for that, never the text. Your AI app's own history (for Claude Code, ~/.claude/projects) is separate and is not touched.

MCP tools

Six tools on purpose: every tool's description is read by the model on every turn, and similar tools get confused with each other. Version 0.1 had thirteen.

See your memory

graph-mind-view              # opens your memory as a 3D graph in the browser
graph-mind-view --no-text    # the same picture with no conversation text in the file

Every conversation turn and memory is a point; the links are the associations recall already uses (shared words and names, neighbours in time). Points that keep coming up together gather into coloured regions, each named by its most particular words. Hover a point to read it; click to fly to it. The page is a local file (~/.graph-mind/brain.html); only the 3D library is fetched from a CDN. Use --no-text before you screenshot or share it.

One memory across PCs

On the PC that holds the memory, tell its AI:

"Let my other PCs use this memory."

It replies with a connection code (gm1.…). On each other PC:

graph-mind-install --join gm1.…     # or: python src/install.py --join gm1.…

The memory then lives in a Postgres server on the first PC, which Graph-MIND sets up itself. Other PCs reach it over the local network in the office, or over Tailscale from anywhere. Each connection tries the addresses in turn and uses the first that answers.

Other PCs log in with a generated password, as a role that can reach only the memory database. The Windows firewall rule admits only the local network and Tailscale. Every line of memory is encrypted (AES with an HMAC) under a key derived from that password before it leaves a PC, so the network and the hub's own database carry only ciphertext. The code contains the password: do not post it publicly.

A synced folder also works. Tell the AI "use my Google Drive's Graph-MIND folder as my memory" on each PC.

Reproducing the benchmarks

1. Download longmemeval_s_cleaned.json from LongMemEval into research/external/longmemeval/.
2. Set OPENAI_API_KEY for the answer model and the judge.
3. pip install -e ., then from research/ run the commands in REPORT.md §10.

A full 500-question run costs about US$3.

Tests

python -m unittest discover -s tests -p "test_*.py"

Known limits

• For the first few seconds after an AI app starts, answers use word search only while the embedding model loads in the background (6-30 s on a slow, synced disk, depending on load; it was 50-70 s before 0.4). The first answer never waits for it; later answers take well under a second.
• Daily use on Windows only so far. macOS and Linux pass the tests in CI but have not seen real use.
• Capture follows each app's transcript format, which is not a public interface. An app update can stop capture until Graph-MIND is updated.
• Multi-session questions are the weakest type at 80%. These are questions that count or combine facts across many conversations.

The full list is in REPORT.md §9.

Repository layout

Security

Graph-MIND holds your conversations. Report vulnerabilities privately (GitHub Security → Report a vulnerability); SECURITY.md lists what is in scope and how the data is handled.

Contributing

Issues and pull requests are welcome. Contributions are accepted under the CLA, which keeps the dual license possible.

License

AGPL-3.0. Anyone running a modified Graph-MIND as a network service, such as the memory behind a support chatbot, must publish that source. A commercial license is available for products that cannot.