15 years shipping software. The last 12 months, nothing but AI — full-time, all in. What came out is 17 tools that make a generic coding agent actually mine; I'm releasing them one at a time, and this is the first. My coding agent woke up amnesiac every session — I re-explained the project, re-decided settled questions, re-found the same bug. So I built it a memory with one rule I don't let it break: when it doesn't know, it says so. It never invents an answer. Recall returns one of three honest outcomes — grounded (evidence with its source), missing_evidence (matched but every match excluded, with per-reason counts), or abstain (nothing matched). There is no fourth "here's my best guess" branch; you can read the source and confirm the code path doesn't exist. Capture is provenance-mandatory: no source, no storage. A disproven fact gets a `falsifies` edge and stops grounding recall — the memory remembers being wrong. It runs entirely on your machine: one Rust binary, one SQLite file, compiled without a networking stack (strace shows zero socket() calls). MCP over stdio, so it plugs into Claude Code / Codex / any MCP client. AGPL-3.0. curl -fsSL https://no.tt/install | sh github.com/menot-you/n-memory I'm one engineer (n=1). Happy to go deep on the design — especially the "advisory, not authority" stamp that makes stored text unable to hijack the agent. #2, nSOUL (the engineering core the rest sit on), is already up; 15 more to come.

nMEMORY

Local memory for AI agents that admits when it doesn't know.

Local memory for AI agents that admits when it doesn't know.