FAQ & limits
Which agents does it work with?
Any MCP client can use the server over stdio (kgl-memory-mcp). The one-command setup (kgl-memory-init, including the Stop hook) targets Claude Code. Other clients need the server registered by hand.
Does my memory leave my machine?
No. The graph is local files, search is local, and nothing is sent anywhere by the server. The optional semantic extra downloads a small model once from Hugging Face. The optional reflect command calls your own claude CLI, which uses your Claude account, and only when you run it.
How big can a memory get?
It's designed for one person's or one project's memory: thousands of nodes. All-pairs operations such as the duplicate scan are quick at that size. It is not a document store for millions of items.
Can several agents or people write at once?
Writes on one machine are serialised with a file lock. Across machines or branches, memory merges like any text: concurrent edits to the same lines can conflict, and you resolve them in git.
Should the memory be in git?
That's the main advantage, so usually yes. If it contains sensitive material, keep the graph outside the repo or in a private one. Consider ignoring .index/.
What if the agent writes something wrong?
Edit the file, or have the agent call delete_memory. For a change of mind rather than a mistake, record the new belief so it supersedes the old one and the history stays readable. Reviewing memory diffs catches these early.
Why doesn't it use a vector database?
Because the goal is memory you can inspect and correct. See Why plain text for the trade-offs, including when you should choose a vector database instead.
What's the licence and status?
MIT-licensed, alpha. It requires Python 3.10+. The format and tools are tested and in real use, and may still change between releases.
What's planned?
Not available yet: starter memory packs, hosted sync and team memory, and MCP directory listings. Want help now? Unfoldata can set it up for your team.