Every morning your AI coding agent wakes up with no idea who you are.
You told it yesterday that the project is on PostgreSQL 16, that payments go through Stripe with idempotent webhooks, that the client hates em-dashes, that the legacy/ folder is off-limits. Today? Blank slate. You re-explain it. Tomorrow you re-explain it again.
That's not a prompt problem. It's a memory problem.
Corvusmem
Corvusmem gives your coding agent (Claude Code, Codex, or a plain CLI) a memory that actually persists — and that never leaves your machine.
- Local. Your memory is a plain file in your repo. No server, no cloud, no API key, no telemetry.
- Deterministic. Same input, same result. No embeddings, no vector database, no "temperature".
- Zero dependencies. It's small enough to read in an afternoon.
- Any language. Keys and values work in English, Spanish, anything — the recall is language-agnostic.
- Shared, per-author. On a team, everyone's facts stay separated and merge cleanly.
30 seconds
# in your project
npx corvusmem install
# remember something
npx corvusmem --save --dimension project.stack \
--value "API is Fastify v5 on Node 22, PostgreSQL 16 + pgvector" \
--keys "stack,backend"
# recall it
npx corvusmem --recall "backend"
Once installed, the agent captures and recalls memory on its own — you don't have to babysit it. There's a hard cap on how much gets injected per message, so your context window stays small and doesn't grow as memory does.
See your memory as a brain
Run npx corvusmem --cortex and it draws your whole memory as a local HTML "brain" — decisions, goals, project facts, each in its own zone. It's offline and it's oddly satisfying.
Honest footnote
Corvusmem is new. I'm not going to wave fake download counts at you. It's open and free for non-commercial use (PolyForm Noncommercial), with a separate commercial license. If your agent forgetting everything drives you up the wall like it did me, give it a spin and tell me where it breaks.
→ Docs & install: https://corvusmem.vercel.app/

Top comments (0)