03 / lucid-mem / Agent memory
Carry context across coding sessions.
Local-first, Git-backed memory for 15 AI coding agent hosts, with deterministic retrieval and a measured save loop.
Bring useful context forward.
At session start, `amr context` assembles a ranked, bounded digest without calling a model in the read path. Agents can save durable knowledge to a Git repository; background reflection can propose missed learnings for review. Maintenance and coverage tools help keep the memory useful and inspect whether important signals were saved.
- Wire one memory repository into 15 agent hosts through hooks where available, and rules or MCP where they are not.
- Rank context by memory type, recency, and recorded use under a configurable size budget.
- Use stable entry IDs, expiry, archive, and Git history to review and correct stored knowledge.
- Run `amr coverage` to replay transcripts, surface missed durable signals, and report a save rate.
- Ranked context at session start
- Agent saves durable signals
- Git history + reviewable maintenance
- Coverage checks for missed signals
Follow the project.
Use the repository for current support, installation, releases, and contribution guidance.