Documentation
A memory your agent keeps.
Corvusmem gives your coding agent a real memory — local, deterministic and offline. It hooks into Claude Code and Codex, writes what your agent learns to a file you own, and recalls it across sessions and projects. No embeddings, no server, no API key.
What it is
Your agent, with a memory of its own — and it never leaves your machine. Agents forget everything between sessions; Corvusmem is the memory that remembers you and your projects, in a single local file you control.
It is deterministic (same input, same result), offline (no network) and zero-dependency. Unlike cloud memories, yours stays yours.
Install
One command. No account, no cloud, no API key — the hook writes to a local file you own.
Works with Claude Code and Codex · the exact command for your host is shown during setup
How it works
Capture
The hook listens
A PostToolUsehook reads each exchange and extracts facts — decisions, file paths, project context.
Store
Facts, not vectors
Each fact lands in a lightweight local graph with its name, date and links. The whole Cortex is a single file you own.
Retrieve
Two-hop activation
A query activates the matching node and spreads one hop out, surfacing related facts — no semantic search, no embedding model. Instant.
The memory
Think of it as a notebook your agent keeps for you — but it only writes down what isstable: decisions, the project's stack, who you are, your preferences, your goals. Passing chatter is ignored. When something changes, the new valuereplaces the old one and the old one is kept, dated, in the history — so your agent is never working from something you already changed.
There are two kinds of memory, and they never get mixed up:
Project
About the project
The stack, the client, decisions, deadlines. It lives in the project folder (.corvusmem/), so your team — and Codex — share it. "The stack is Next" is true no matter who said it.
You
About you
Who you are and how you like to work — carried across every project. It lives on your machine (~/.config/corvusmem/), private to you, never committed. Your agent recalls both at once.
Teams — one project, many people
When several people work on the same project, the shared memory keeps each person'spersonal facts separate. If Ana prefers Vim and Bob prefers Emacs, both are kept — one never overwrites the other. Facts that describe theproject itself (stack, decisions) stay shared and have no owner, exactly as they should.
You don't configure anything: Corvusmem reads the author from your gitemail (or your system user), with no network. For a solo developer it's invisible — it just works.
The Cortex
The Cortex draws everything your agent knows as a living brain of points — one glow per fact. Each kind of memory settles in its own zone: decisions, goals, project facts, who you are, what was lived.
Search a fact and only its node and connections light up, like a neuron firing. It is not a picture — it is a readout of the memory your agent actually uses.
Open your own Cortex from any project folder — no extra setup:
It writes a self-contained HTML file (offline, no network) and prints the path — open it in your browser.
Hosts
The engine detects where it runs and adapts; the memory is the same underneath.
Claude Code
Hooks register in your project settings; facts captured after every tool call.
Codex
Runs through the plugin path; the same brain, the same local file.
Standalone
A plain CLI for any project — read, write and search the memory directly.
License
Free for personal and non-commercial use underPolyForm Noncommercial. Zero telemetry — your memory lives in local files you own.
Using Corvusmem inside a company or a paid product needs a commercial license.
FAQ
Does my data ever leave my machine?
No. The memory is a local file you own. No server, no cloud, no API key, no telemetry. Corvusmem runs entirely offline.
How is this different from mem0 or Zep?
Those store your memory in their cloud behind an API. Corvusmem keeps everything local and deterministic — same input, same result, no embeddings or vector database involved.
Do I need an embedding model or a vector DB?
No. Facts are stored in a lightweight local graph and retrieved by name, date and relevance with two-hop activation. No embeddings, no vector database.
Which agents does it work with?
Claude Code and Codex today, plus a standalone CLI. The engine detects its host and adapts; the memory is the same underneath.
What happens with a team on the same project?
Shared project facts (stack, decisions) are common to everyone. Personal facts (your preferences, who you are) are kept per author, so two teammates never overwrite each other. The author is read from your git email, with no setup and no network.
Is it free?
Free for personal and non-commercial use under PolyForm Noncommercial. Using it inside a company or a paid product needs a commercial license.