A local, provider-neutral MCP server for scoped and reviewable preference memory.
Strong schema definitions with proper type constraints and comprehensive parameter documentation. All 6 tools have clear descriptions (50-150 chars) and well-structured input schemas using Pydantic annotations. Tool names follow verb_noun convention (memory_recall, memory_observe, etc.). Tool annotations are properly declared (READ_ONLY, SAFE_WRITE, STATE_CHANGE). However, output schemas are not explicitly documented in the source, return types are inferred from model classes but not visible in tool registration. Error handling is minimal; no recovery guidance or actionable error messages visible. Missing dependency hints and multi-step guidance.
Read one memory by its opaque identifier.
Use host sampling to extract pending candidates without storing the raw observation.
Recall active memories that match the current request and opaque scope selectors.
Create a pending, structured memory proposal without invoking sampling.
Approve, reject, or dispute a memory after a real form-elicitation response.
List reviewable memories with cursor-based pagination.
Output schemas not documented. Return types (RecallResult, ObserveResult, etc.) are defined in models but not visible in tool registration. LLMs cannot plan downstream operations without knowing response structure.
No error recovery guidance. Tools lack descriptions of failure modes, retryability, or next steps. E.g., memory_resolve does not document what happens if memory_id is invalid or if supersedes_id references a non-existent memory.
memory_observe and memory_remember descriptions lack dependency hints. They reference 'host sampling' and 'pending' states but do not explain when to call memory_review or memory_resolve to finalize memories.
Inferred effective spec: <=2025-11-25.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | D | 59 | <=2025-11-25 | v2 |
memory_get description is generic ('Read one memory by its opaque identifier'). Does not explain when to use it vs memory_recall, or what happens if the memory_id is invalid.
Sampling integration is server-initiated (deprecated pattern). memory_observe and memory_recall conditionally invoke ctx.session.create_message() for sampling. Should integrate directly with LLM provider API or document as optional enrichment.