Graph-MIND

Local-first memory for AI coding assistants. Verbatim storage, on your own PC: 88.8% on LongMemEval with zero model calls at write time.

Switch models; keep the memory.

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

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).
• Nothing leaves your machines. There is no cloud, no account and no telemetry.

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 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: same 40 questions, same answer model, prompt and judge. The run was pre-registered with code hashes; nobody had tuned on these questions.

Graph-MIND's lead over both is significant (exact McNemar p = 0.002 and 0.003).

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.

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 WinError 1114 loading c10.dll, Windows is missing the Microsoft Visual C++ runtime that PyTorch 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 install.py

The installer:

• installs the packages (PyTorch CPU is the large one, about 2 GB);
• registers the MCP server with every app it finds: Claude Code, Claude Desktop (including the Microsoft Store build) and Codex;
• starts the capture service at login (Windows Startup folder, a macOS LaunchAgent, or an XDG autostart entry on Linux);
• downloads the embedding model.

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.

What it captures

Before anything is stored, these are masked: API keys, tokens from GitHub, AWS, Google and Slack, private keys, and passwords inside URLs.

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:

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

It deletes every captured turn and memory containing the phrase from this PC and its indexes, removes their lines from the shared memory, and tells your other PCs to delete their copies on their next sync. Only ids are shared for that, never the phrase. Your AI app's own history (for Claude Code, ~/.claude/projects) is separate and is not touched.

MCP tools

Five 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.

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 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. 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 external/longmemeval/.
2. Set OPENAI_API_KEY for the answer model and the judge.
3. Run the commands in REPORT.md §10.

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

Tests

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

Known limits

• The first question after an AI app starts waits for the embedding model to load (a few seconds; 30-50 s on a slow or synced disk). Later questions 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 repository still holds the research-phase experiments next to the product (see Repository layout); the installed package carries only the 21 product modules.

The full list is in REPORT.md §9.

Repository layout

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.