MCP memory server with spaced repetition freshness tracking, backed by Elasticsearch
This server has significant definition quality gaps. While tool names are clear (verb-based: inspect_, create_, update_, delete_, read_, list_), descriptions vary widely in quality and completeness. Most critically, many parameters lack descriptions entirely, and output schemas are completely undocumented. The server has 11 tools operating on a knowledge graph backend, but the definition quality does not meet production standards. Parameter documentation is sparse (e.g., keywords array in inspect_files is required but the description doesn't explain why or what makes a 'good' keyword). Output schemas for knowledge graph operations are missing entirely, LLMs cannot reason about what results look like. Error handling guidance is absent from all tool descriptions. Tool compositions are reasonable (inspect_files, inspect_knowledge_graph, and CRUD operations on entities/relations form coherent chains), but the lack of output documentation breaks downstream planning.
Create entities in knowledge graph (memory). Do NOT store information derivable from source code, git history, or documentation — memory is for context, decisions, and knowledge not in the codebase. Set reviewInterval based on content volatility: stable facts (180-365 days), project state (14-30), volatile state (1-7).
Create a new memory zone for organizing entities
Create relations between entities in knowledge graph (memory)
Delete entities from knowledge graph (memory)
Delete a memory zone and all entities within it
Delete relations between entities in knowledge graph (memory)
Agent driven file inspection that uses AI to retrieve relevant content from multiple files.
Output schemas are completely undocumented. LLMs cannot reason about what fields are returned from inspect_files, inspect_knowledge_graph, read_graph, or any other tool. This violates the 'Document the output schema' critical check and makes tool composition planning impossible.
Many tool descriptions are too brief or generic. 'Delete entities from knowledge graph (memory)' (45 chars) is under the 10-1024 character guideline but lacks context on consequences, prerequisites, or recovery. 'Update entities in knowledge graph (memory)' does not explain when to use it vs create_entities or how the update semantics work (merge? replace?).
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-22 | F | 46 | 2026-07-28+ | v2 |
| 2026-03-09 | F | 0 | - | v1 |
Agent driven knowledge graph inspection that uses AI to retrieve relevant entities and relations. Results include freshness metadata (confidence, needsReview). Treat memory as 'what was believed true at a point in time.' Trust current observations (code, logs, git) over recalled memory when they conflict.
List all available memory zones
Read the entire knowledge graph or a specific memory zone
Update entities in knowledge graph (memory)
Critical parameter descriptions are missing. In inspect_files, the 'keywords' parameter is marked REQUIRED but the description does not explain why keywords are essential or what makes them effective (e.g., 'AI will use these to filter files, be specific, use synonyms, include domain terms'). In update_entities, the 'observations' array in entities has no description of what constitutes a valid observation or how updates are merged with existing ones.
No error handling or recovery guidance. None of the tool descriptions tell LLMs what to do if a memory_zone is not found, if an entity name is empty, or if a deletion succeeds but related relations remain orphaned. The source code shows tests for these cases (test-empty-name.js, test-non-existent-entity.js, test-relationship-cleanup.js), but the tool definitions offer no guidance on error classification or recovery steps.
read_graph has an ambiguous description ('Read the entire knowledge graph or a specific memory zone'). The output structure is not documented, does it return {entities: [], relations: []}? Nested zones? Does it include metadata like confidence scores (mentioned in inspect_knowledge_graph description) or review intervals (mentioned in create_entities)?
Inconsistent parameter naming and documentation. create_entities takes 'memory_zone' and 'entities' array with nested 'reviewInterval' field, but other tools use 'zone' (delete_memory_zone). Parameters like 'include_entities' and 'include_relations' in inspect_knowledge_graph are boolean flags but lack description of what 'full details' vs 'relations' means in response format.
Destructive operations (delete_entities, delete_relations, delete_memory_zone) lack confirmation or dry-run support. Descriptions do not warn LLMs of irreversibility or suggest a cautious approach. No anti-pattern guidance prevents agents from bulk-deleting critical memory.
inspect_knowledge_graph description mentions 'freshness metadata (confidence, needsReview)' but input parameters have no way to filter by freshness, confidence, or review status. The response structure is not documented, so LLMs cannot plan how to use this metadata.