Why Memmo
Your agents already have three kinds of memory. None of them is the team's.
Comments explain code. CLAUDE.md tells one agent the rules. Claude's own memory remembers you. Memmo is the memory of the team: captured from real work, reviewed by people, and served to every agent at the moment it's relevant.
| Which memory | Code comments | CLAUDE.md | Claude's memory | Memmo |
|---|---|---|---|---|
| Shared by the team | ✓in git | ✓in git | –per user | ✓per project |
| Spans several repos | –one repo | –one repo | ~not repo-aware | ✓project-wide or per repo |
| Works in every agent | ~where read | –Claude Code | –one tool | ✓any MCP agent |
| Served for the task | –found by chance | ~whole file | –opaque | ✓per task & file |
| Captured automatically | –by hand | –by hand | ~unreviewed | ✓sessions, PRs |
| Reviewed by people | – | ~if code-reviewed | – | ✓inbox |
| Typed and scoped | –prose | –prose | –prose | ✓rule, risk, … |
| Improves with use | –rots | –rots | –unknown | ✓feedback loop |
| Auditable per session | – | – | – | ✓live feed |
✓ yes · ~ partly · – no. The details are below.
Code comments
They explain a line, to whoever reads that file
A comment is attached to a place in the code, written at the moment the code was. It can't hold what has no place: “we tried Redis here and rolled it back”, “anything under billing needs the payments team”, “never seed the production database”. Nobody retrieves comments; they are found by accident, and nobody knows which ones are stale.
CLAUDE.md
A page of rules, read whole, every session
It's the right tool for a page of stable conventions, and Memmo writes its own section into it on purpose: it's the entry point, not the competitor. It stops working as it grows. The whole file competes with your task for context, it's the same text whether you're editing a migration or a React component, and someone has to remember to update it. It only speaks to Claude Code.
Claude's own memory
Remembers you, not your team
What one agent learns in one tool stays there. Your teammate's Claude doesn't know it, Cursor doesn't know it, and you can't audit or approve what was kept. It's written from the model's judgement of what mattered, with no review step, and it doesn't come back scoped to the file you're about to change.
Memmo
The team's memory, served when it matters
- ✓Shared across the team, across repos, across agents. One project holds the memory of a monorepo or of several services; each memory applies to every repo or to the ones you pick. Claude Code, Claude Desktop, Codex, Cursor or anything speaking MCP. A colleague's session in the API repo yesterday teaches your session in the web repo today.
- ✓Retrieved, not dumped. A short briefing at session start, then recall per task, and risk warnings only for the files being touched.
- ✓Captured automatically, approved by humans. Sessions, scans and merged pull requests propose memories. The Inbox is where a person accepts, edits or rejects them. Corrections supersede old facts instead of piling up contradictions.
- ✓Typed and living. Rule, risk, decision, failure, command. Each with confidence, usage counts and evidence. Reviewer feedback raises what helps and fades what doesn't.
- ✓Visible. A live feed shows what was injected into which session, and by whom.
Where to start
Keep your CLAUDE.md. Memmo grows out of it.
Memmo doesn't replace the file you already have: memmo init writes its own section into it and takes over from there. The day a second person, a second repo or a second tool shows up, the memory is already shared, already reviewed, and already served where it's needed. Nothing to migrate, nothing to rewrite.
It stays out of the way, too. Hooks are silent and fail open, so an outage or an unconnected repo never blocks a session. And nothing reaches an agent without a person approving it first.