Unified memory and skill infrastructure for AI agents. Provides MCP server exposing search, retrieval, and memory management over a markdown vault with SQLite FTS5 + sqlite-vec hybrid search.
Memstem demonstrates solid tool design with clear naming (verb_noun pattern), comprehensive parameter schemas, and well-structured output. All 5 tools have descriptions and input schemas. However, parameter descriptions vary in quality, some are detailed (memstem_search) while others are minimal (memstem_list_skills has empty input). Error handling guidance is absent from descriptions. Output schemas are documented in code but not exposed to LLMs via tool definitions. The server follows good composition patterns (single responsibility per tool) and includes natural identifiers (id_or_path). Main gaps: no error recovery guidance in descriptions, no output schema documentation in tool definitions, and missing parameter constraints (e.g., importance_weight range, mmr_lambda bounds).
Retrieve a single memory by ID or path. Returns full memory body, metadata, and frontmatter.
Retrieve a single skill by ID or path. Returns full skill body and metadata.
List all skills in the vault. Returns metadata for each skill record without full body content.
Search the vault using hybrid BM25 + vector search with optional reranking. Returns ranked results with snippets, scores, and metadata.
Create or update a memory in the vault. Accepts markdown body with optional YAML frontmatter. Returns the created/updated memory with assigned ID and metadata.
memstem_list_skills has empty input schema (no parameters documented) and minimal description. LLMs cannot determine what this tool filters or returns without calling it.
Parameter constraints missing for numeric fields: importance_weight, bm25_weight, vector_weight, mmr_lambda lack min/max bounds. LLMs may pass invalid values (e.g., importance_weight > 1).
No error recovery guidance in tool descriptions. If search fails or memory not found, descriptions do not tell LLMs what to do next (e.g., 'Try memstem_list_skills to discover available types').
Output schemas (SearchHit, MemoryDetail) are defined in http_server.py but not exposed in MCP tool definitions. LLMs cannot see what fields to expect from results.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | C | 66 | 2026-07-28+ | v2 |
memstem_upsert accepts optional 'id' to update existing memory but description does not clarify idempotency semantics. Agents need to know: is this call safe to retry? Does passing the same ID twice cause issues?