Your agents learn, recall, reinforce — and forget what stops working. One memory, several agents: whichever one wrote a skill, both can use it.
Two independent reviewers read the whole codebase and then re-read every fix, forty rounds deep, each round hunting for what the previous one broke. What they found was not in the recap hook everyone stared at: the HTTP server let one agent read another agent's private titles through conflict warnings, backlinks and crowded pages; trust could be granted by any process; a dump restore stripped approvals it should have kept. 0.11.0 closes all of it — on HTTP, MCP, CLI and the hooks alike — and every claim in the notes was tried with a command before it was written. Release notes · changelog · known limits
No LLM calls to store or retrieve a skill. Fully local — your memory never leaves the machine, and it costs $0 to use.
Lexical FTS plus a local semantic layer. Write in one language, recall in another — it just finds the skill. The cross-language part is the semantic layer's doing, so install it: pip install 'skillmem[semantic]'.
Skills that help get stronger. Skills that sit unused fade out — the memory curates itself, like yours does.
Every change is chained with SHA-256. You can prove what your agent learned, when, and that nobody rewrote it.
Every memory records where it came from. A rule an agent saved from a README it was reading is not the same as a rule you set — so only you approve one, and everything else arrives marked as data, not instructions.
hit@5 0.871 on the full LongMemEval oracle set — hybrid retrieval, k=5, CPU only, reproducible from the repo. We publish the retrieval mode and the embedding model with the number.
The full engine — learn, recall, reinforce, decay — is open source and on PyPI. pip install 'skillmem[semantic]', then skillmem init --claude-code --codex.
Join the waitlist — early access goes out in order of signup.
One email when cloud opens. No spam, no sharing.