Experimental local-first memory governance for AI agents.
Mimir is built around one rule: agents may propose memory, but they do not write trusted shared memory directly. Session notes, checkpoints, native agent memories, and adapter exports enter as untrusted drafts. The librarian validates, structures, deduplicates, and commits accepted records into an append-only canonical log with provenance.
The project is pre-1.0. The architecture and local implementation are real, but storage details, CLI flags, draft envelopes, and adapter workflows may change before the first stable release.
| Area | Current state |
|---|---|
| Append-only store | Canonical log, replay, verification, crash recovery, symbol tracking, supersession, and confidence decay are implemented in mimir-core. |
| Librarian path | mimir-librarian ingests draft envelopes, validates candidate canonical records, filters duplicates, handles conflicts, and commits through the governed store path. |
| Agent harness | mimir <agent> [agent args...] preserves the native terminal flow for local agents while adding bootstrap, context, checkpoint, capture, and post-session handoff hooks. |
| Operator controls | mimir status, mimir doctor, mimir context, mimir drafts ..., and mimir memory ... expose setup, bounded context, draft triage, and read-only memory inspection. |
| MCP | mimir-mcp exposes governed local memory tools over stdio MCP. |
| Recovery | Git-backed `mimir remote status |
| Benchmarks | Recovery benchmark fixtures, launch contracts, transcript gates, and score validation live under benchmarks/recovery. Public benchmark claims wait for recorded live runs. |
| Codex | plugins/mimir is a coherent Codex plugin bundle for the Mimir workflow. It is not a standalone memory skill and does not bypass the librarian. |
- Production readiness.
- Stable storage, CLI, API, MCP schema, or wire-format compatibility.
- Hosted service availability.
- Benchmark-proven superiority over other memory systems.
- Direct agent writes into canonical memory.
- Cross-project or operator-wide memory promotion without librarian governance.
git clone https://github.com/buildepicshit/Mimir.git
cd Mimir
cargo build --workspace
cargo test --workspace
cargo run -p mimir-harness -- doctor --project-root .
cargo run -p mimir-harness -- rustc --versionFor a fresh-clone walkthrough, see docs/first-run.md.
Build the transparent harness:
cargo install --locked --path crates/mimir-harnessInspect project readiness:
mimir doctor --project-root .
mimir status --project-root .Wrap an agent session:
mimir codex
mimir claudeRecord an intentional draft memory from a wrapped session:
mimir checkpoint --title "short title" "memory note"Process captured drafts after a session with the configured per-repo librarian policy. The canonical write boundary stays the librarian path; raw native memories and checkpoint notes are draft evidence until accepted.
Mimir's recovery path is explicit Git-backed BC/DR mirroring:
mimir remote status
mimir remote push --dry-run
mimir remote push
./scripts/bcdr-drill.sh --dry-runmimir remote push mirrors the append-only workspace log and draft JSON files into the configured recovery repository. mimir remote pull restores missing or prefix-safe local state after verifying canonical-log integrity. See docs/bc-dr-restore.md.
docs/README.md- public documentation index.STATUS.md- current implementation and release state.PRINCIPLES.md- engineering principles and architectural boundaries.docs/concepts/- architecture specs.docs/launch-readiness.md- OSS, engineering, and promise sign-off checklist.RELEASING.md- release process and package workflow.docs/blog/2026-04-28-agent-memory-compiler-pipeline.md- launch article.
Useful review areas are Rust correctness, append-only log integrity, security boundaries, recovery benchmark methodology, adapter UX, and documentation clarity.
See CONTRIBUTING.md. All contributors follow CODE_OF_CONDUCT.md. Report security issues per SECURITY.md, not as public issues.
Apache-2.0. See LICENSE.
Mimir is a BES Studios project.