Fully local semantic memory for AI agents: search your notes and session logs from any MCP client, with embeddings served by Ollama
Single tool 'search_memory' has a well-crafted description (169 chars, within 10-1024 baseline) and proper verb-noun naming. Input schema is complete with typed parameters (query: string, top_k: integer, section_filter: string). However, output schema is undocumented, the tool returns TextContent but the LLM cannot see what fields or structure to expect from the search results. Error handling is present (Ollama connection, timeout, ValueError) but lacks recovery guidance per pattern:recovery-guide. No parameter constraints (enums, min/max) despite section_filter having known valid values. Missing documentation of output structure, pagination behavior, and result limits.
Semantic search over your local notes and session logs (fully local: embeddings via Ollama, nothing leaves your machine). Use it before design, implementation, or debugging work to recall past decisions, solutions, and failure patterns.
Output schema undocumented. Tool returns TextContent but LLM cannot see result structure (fields, format, pagination). Per pattern:tool and pattern:response-shaper, structured output with typed fields is required.
section_filter parameter accepts enum values (derived from section_types) but schema does not enforce enum constraint in all code paths. When section_types is empty, filter_schema lacks enum, LLM cannot discover valid options.
Error messages lack recovery guidance. Ollama timeout returns 'Embedding timed out...' but does not suggest next steps (retry, increase timeout, check Ollama status). Per pattern:recovery-guide, errors must tell LLM what to do.
top_k parameter lacks min/max constraints. Unbounded integer could allow LLM to pass absurd values (0, 1000000) that break performance or memory.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | C | 61 | 2026-07-28+ | v2 |
No documentation of result limit or pagination. If index contains 10,000 chunks, does top_k=5 always return 5? What if fewer matches exist? LLM needs to know expected result cardinality and behavior.