| name | supamem |
| description | Project-agnostic dual-memory tooling — semantic + structural memory for AI coding agents |
supamem skill
supamem packages a battle-tested semantic memory layer (Qdrant hybrid sparse+dense
retrieval), an MCP server, per-client hooks, and a usage counter as a single CLI.
When to use which subcommand
| Subcommand | Use when |
|---|
supamem init | First time on a fresh project — probe Qdrant, create collection, write .supamem/config.toml. |
supamem index | After updating project docs, ADRs, insights, or rules. Re-embeds changed docs only. |
supamem index --snapshot cursor | Refresh .cursor/rules/dual-memory-snapshot.mdc (Cursor's passive channel). |
supamem mcp-server | Run the MCP server (clients usually call this themselves; manual invocation only for debugging). |
supamem hook claude-code --file-path X | Per-edit semantic-context injection for Claude Code (clients call this). |
supamem doctor | Probe Qdrant health, resolve config chain, report version drift. Run when something feels off. |
supamem stats --show today | See today's tool-call rate, p95 latency, and which client is calling. |
supamem migrate | Brownfield: convert a legacy dev_memory collection into the supamem schema. |
supamem install --client X | Wire supamem into Claude Code / Cursor / OpenCode. Idempotent. |
supamem uninstall --client X | Cleanly reverse supamem install. |
Typical first-time setup
uv tool install supamem
docker run -d -p 6333:6333 -v $HOME/.qdrant:/qdrant/storage qdrant/qdrant:latest
cd <your-project>
supamem init --yes
supamem install --client claude-code
supamem doctor
Notes
- Locked schema (Phase 80.1): hybrid BM25 + MiniLM with RRF fusion, T-1 markdown chunker.
- Fail-soft contract: hooks and counter never block the calling tool — observability before correctness.
- MCP tool exposed by
mcp-server is named dual_memory_search.