MCP server — persistent LLM memory backed by SQLite
Server has 8 tools with complete input schemas and descriptions. Naming is action-verb based (add, update, delete, get, list, search, topics, dreaming) and mostly clear. Descriptions are present but vary in quality, some are terse (topics: 76 chars), others include workflow guidance (list, dreaming). All tools have JSON Schema with type definitions and constraints (e.g., importance 1-5, limit bounds). However, parameter descriptions are minimal or missing context (e.g., 'Memory ID to update' lacks guidance on what happens if ID doesn't exist). Output schemas are not formally documented, responses are inferred from handler code. Error handling is basic (throws on missing ID) but lacks recovery guidance. No tool annotations (readOnlyHint, destructiveHint) despite clear risk levels. Duplicate detection in 'add' is undocumented. The 'dreaming' tool includes embedded workflow instructions, which is creative but not standard MCP practice.
Store a new memory. Checks for duplicates in the same topic first.
Delete a memory by ID.
Get memories for cleanup and reorganization.
Fetch a single memory by ID.
List memories. SESSION START: call list(topic="_bootstrap",limit=1) first — contains identity, workflow rules, and memory_map with exact record IDs.
Full-text search across content, topic and keywords.
List all topics. Fallback only — prefer list(topic="_bootstrap",limit=1) at session start.
Output schemas not formally documented. Responses are inferred from handler code (e.g., dreaming returns {success, total, memories, has_more, next_offset, instructions}). LLMs cannot plan downstream calls without knowing return structure.
Tool annotations missing. 'delete' is destructive, 'update' and 'add' are writes, but no destructiveHint or readOnlyHint tags present. Agents cannot distinguish safe from risky operations.
Error handling lacks recovery guidance. 'Memory #123 not found' is thrown but does not suggest next steps (e.g., 'Try list() to see available IDs' or 'Try search() to find by content'). Agents cannot self-correct.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | D | 53 | 2026-07-28+ | v2 |
Update content, keywords or importance of a memory by ID.
Parameter descriptions are minimal. 'Memory ID to update' does not explain what happens if the ID is invalid, whether partial updates are allowed, or what fields can be null. Ambiguity invites misuse.
Duplicate detection in 'add' is undocumented. Handler checks for similar entries in topic but description does not mention this behavior. LLMs may not expect rejection and retry with identical input.