This release turns mcp-brain from a local MCP memory layer into a benchmarked, measurable awareness system for Claude Code and MCP-compatible coding agents.
Highlights
- SWE-bench Full benchmark: evaluated on 2,294 real GitHub bug-fix tasks.
- 63.4% Hit@10 file localization with BM25 + code graph + optional semantic reranking.
- Zero LLM cost for retrieval and file prediction.
- Sub-100ms prediction pipeline in the default local-first flow.
- Token-efficient context loading: compressed project awareness in ~100 tokens instead of repeated repo exploration.
- Lifecycle-aware memory: active, suspect, stale, and superseded decisions.
- Team-aware coordination: soft claims, conflict detection, and ownership tracking.
- Improved documentation with architecture, memory hierarchy, prediction pipeline, benchmark results, and token-efficiency diagrams.
Why this matters
Claude Code does not need more raw context. It needs structured awareness of the repo, the team, and past decisions.
mcp-brain v1.4.0 provides that awareness locally through SQLite, BM25, code graph expansion, optional semantic reranking, and MCP tools that Claude can call when starting a session or working on a ticket.
Benchmark summary
| Metric | Top-1 | Top-3 | Top-5 | Top-10 |
|---|---|---|---|---|
| Hit | 24.5% | 43.4% | 53.7% | 63.4% |
| Recall | 20.1% | 36.6% | 46.1% | 55.8% |
| MAP | 24.5% | 28.4% | 30.4% | 31.8% |
Evaluated on SWE-bench Full with 2,294 instances, test files excluded by default.
Recommended install
pip install -e ".[all]"