Self-improving MCP memory server for AI agents. One command, no cloud, no config.
Lorekeeper demonstrates solid tool design with complete input schemas and descriptions for all 10 tools. Naming follows verb_noun conventions (lore_search, lore_insert, lore_remember). Descriptions are substantive (100-250 chars typical), explaining WHAT the tool does and WHEN to use it. However, several tools lack output schema documentation, parameter descriptions are sometimes generic, and error handling guidance is absent. The semantic memory domain is well-served, but composition patterns could be stronger, some tools (lore_insert, lore_update) accept generic object arrays without field-level schema clarity.
Soft-delete one or more memories by ID.
Retrieve pending link suggestions for review, sorted by quality score.
Insert memories and/or links into the store.
Return all session IDs that have been marked as processed via lore_reflect.
Suggest link candidates between a memory and related memories. Returns ranked candidates with per-signal scores. Does NOT write any links — call lore_insert with links=[] to confirm.
Reflect on a completed session — save what you learned.
Capture a thought instantly — one fact, one call.
lore_insert and lore_update accept generic object arrays (memories, links, feedback) without documented field schemas. LLMs cannot infer required fields, types, or constraints for nested objects.
No output schemas documented for any tool. LLMs cannot plan downstream calls or extract required fields (e.g., does lore_search return memory_id, lore_id, or id?). Breaks tool chaining.
No error handling guidance. Tools lack recovery hints (e.g., 'If memory not found, try lore_search() first'). Agents cannot self-correct on failures.
Inferred effective spec: <=2025-11-25.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | C | 62 | <=2025-11-25 | v2 |
Accept or reject one or more link suggestions in a single call.
Search memories by semantic + keyword query, or bulk-fetch by ID. When ``ids`` is provided, skips the vector/BM25 pipeline entirely and does a direct SQL lookup by lore_id. ``query`` is ignored in that path.
Rate memories and links after using them. Drives the quality signal loop.
lore_insert and lore_update lack field-level descriptions for nested object parameters. 'memory objects' and 'feedback objects' are vague, what fields are required? What are valid values?
lore_reflect's auto_insert parameter defaults to true without warning. If an agent calls this without explicitly setting auto_insert=false, memories are created automatically, potential for unintended side effects.