Corvusmem: a memory for your AI coding agent that never leaves your machine
Your AI coding agent is brilliant for about one session. Then it forgets your stack, your decisions, who you are, and the three things you told it yesterday. Every new chat starts from zero.
I got tired of re-explaining myself to Claude Code and Codex, so I built them a memory: Corvusmem.
What it is
Corvusmem is a deterministic, offline, zero-dependency memory for AI coding agents. It hooks into where you already work (Claude Code, Codex, or a standalone CLI), writes down the stable facts your agent learns, and recalls them across sessions and projects.
The whole memory is a plain file you own. No server, no cloud, no API key, no telemetry, no vector database. Same input, same result — every time.
npx corvusmem install --claude That's it. The hook starts capturing from the next message.
How it works (no embeddings, no magic)
- Capture. A
PostToolUsehook reads each exchange and extracts facts — decisions, file paths, your stack, your preferences, goals. - Store. Each fact lands in a lightweight local graph with its name, date and links. Not embeddings. Not a vector DB. A file.
- Recall. A query activates the matching node and spreads one hop out (two-hop activation), surfacing the related facts. Instant, and the same answer every time.
Because it's deterministic and local, there's nothing to debug in the cloud, nothing to pay for, and nothing leaves your laptop.
It remembers you, and your teammates
Facts come in two kinds:
- Project memory — stack, client, decisions. Lives in the repo, shared by the team.
- Your memory — who you are and how you like to work, carried across every project, private to your machine.
And in a shared project, Corvusmem keeps per-author facts: if you prefer Vim and your teammate prefers Emacs, both survive — neither overwrites the other.
See it: the Cortex
Run npx corvusmem --cortex and your memory renders as a brain of points — one glow per fact, each kind of memory in its own zone (goals, project, rules, who you are, what was lived). It's not a decoration; it's a readout of what your agent actually uses. Local HTML, offline, and it follows your light/dark theme.
The honest part
Corvusmem is not the highest-recall memory on the public benchmarks. On LoCoMo it's competitive (~25% overall) and trades blows with embedding-based memories — it even wins on knowing when to say "I don't know" (abstention) instead of hallucinating.
Where it actually wins is the stuff a benchmark doesn't score: it's deterministic, offline, free of dependencies and API keys, costs nothing to run, and the data is yours. For a coding agent's working memory, I'll take "boring, private, and reproducible" over "a few points higher but in someone's cloud."
Why a crow?
Crows are the memory animal — they remember faces for years and recall thousands of cached locations. Corvusmem = corvus (crow) + mem (memory). The bird that doesn't forget.
Free, and fair
Corvusmem is free for personal and noncommercial use under PolyForm Noncommercial. Using it inside a company or a paid product needs a commercial license. Zero telemetry, always.
Try it
npx corvusmem install --claude - 🌐 Site: https://corvusmem.vercel.app
- 📦 npm: https://www.npmjs.com/package/corvusmem
- 🐙 GitHub: https://github.com/petersonsenadevs/corvusmem
If you try it, I'd love to hear what your agent remembered that surprised you. And if it helped you — a ❤️ on this post means a lot, and any improvement or idea is more than welcome. Drop a comment or open an issue. 🐦⬛