damem

damem tells agents how to manage your project memory and skills files in the code repository without branding them.

Memory of your agents such as Claude Code or Codex lives on one machine in one tool's format, so a new computer, a teammate's clone, or a different agent starts from nothing. .agents/ is committed with the code, and whoever shows up next reads it.

This tiny CLI tool aims to be platform-agnostic and agent-agnostic.

Quickstart

Add the lines below to the top of your AGENTS.md

At the start of every new session, run `damem recall` in the terminal and treat
its output as the context for this repository. If the `damem` command is not
found, read [INSTALL.md](https://github.com/cunarist/damem/blob/main/INSTALL.md)
and follow it before reasoning or responding.

What it does

.agents/ holds plain Markdown that any coding agent can read: what the project knows, what it can do, and where to put scratch files.

damem prints the rules for keeping it, and reports when the directory drifts. It never owns the files, and never writes one. Delete damem and everything there still works.

The rules ship with the binary, so each repository holds only the short pointer above, and upgrading damem updates every repository at once.

.agents/
  memory/
    db-choice.md   one fact per file
  skills/
    release-staging/
      SKILL.md     when to use the skill and how
      upload.py    helper files live with the skill
  scratch/
    .gitignore     scratch space, already ignored

There is no index to keep in sync. Every file describes itself, and damem recall collects those descriptions, so one command shows the agent everything the project has.

Notes

• damem recall tells agents that .agents/ is theirs to edit and that everything outside it needs your approval, so a session that learns something writes it down instead of rewriting your code.
• Skill scripts are Python 3.12 or newer so they run on any operating system. That is a convention damem recall states, not something damem enforces.
• Output is colored for a terminal and plain everywhere else, so damem recall > context.md and piping into an agent both give clean Markdown.
• Reading is sequential and stays out of the way: 500 memories take about 110 ms, 2000 take about 300 ms.