← sources

Claude Code kept forgetting decisions after compaction, so I built a memory it can't overwrite without me

source: reddit (reddit:ClaudeCode)published 2026-10-20 17:40ingested 2026-10-21 02:40

retained text

Long Claude Code sessions kept hitting the same wall for me. After a compaction, or in a fresh session the next day, the "why" behind earlier decisions was gone. Claude would re-litigate settled choices or quietly break a constraint we'd agreed on. CLAUDE.md helped, but it doesn't say which decisions were actually confirmed.

So I built MUON, a governed harness that sits above coding agents.

How it works:
- Agent claims start as proposals, pass deterministic checks, and become confirmed only when I confirm them. Decisions survive compactions and new sessions.
- Three graphs over one ledger: a code graph (so an agent sees an edit's blast radius), a memory graph and a coordination graph.
- Role contracts: each agent gets owned files, allowed tools, a sandbox, required evidence and a budget. Roles can narrow permissions, never widen them.
- A person approves merges, releases, extra spend, widened permissions and secrets access.
- The brain runs on loopback only, and MUON never holds your vendor token.

It's free for individuals. It also works with Codex, Cursor and OpenCode if you mix agents. https://getmuon.com (there's a "watch a mission run" replay)

I'm the builder, so I'm biased. What do you do today to keep decisions consistent across long sessions?

document fetched 2026-10-21 00:39 · canonical https://reddit.com/r/ClaudeCode/comments/1xayqza/claude_code_kept_forgetting_decisions_after