Local-first, project-aware durable memory through a versioned MCP interface
This server demonstrates solid fundamentals with 10 well-named tools and comprehensive parameter schemas. All tools have descriptions, though many are moderately detailed. The tool naming follows verb_noun conventions (list_, read_, create_, update_, delete_, search_), which is excellent. Input schemas are present and typed for all tools. However, several issues prevent a higher score: (1) output schemas are not documented in the tool definitions, (2) no error handling guidance in descriptions, (3) descriptions lack actionable context about when to use each tool vs alternatives, (4) no documented return types or response formats, (5) no enum constraints on status/type/scope parameters despite them accepting 'optional filtering', (6) missing idempotency guidance for write operations, (7) no pagination guidance despite list_memories potentially returning large result sets.
Register a new project by name and optional path
Create a new memory with title, body, and optional metadata
Soft-delete a memory by ID (marks as deleted without permanent removal)
Get the current status of the memory system including document count, indexing state, and health metrics
List all memories with optional filtering by project, status, type, scope, or search query
List all detected or registered projects
Read a memory by ID with full content and metadata
Trigger a full reindex of all memories in the vault
Output schemas not documented. Users cannot see what fields list_memories, search_memories, or get_status return, forcing them to infer structure from trial and error. This violates pattern:tool requirement that all tools document output types.
Enum constraints missing on categorical parameters. 'status', 'type', and 'scope' parameters are described as 'optional filtering' but no valid values are enumerated. This forces LLMs to hallucinate values. Examples: list_memories accepts status but valid options (active, deleted, archived) are not specified.
Error handling not documented. create_memory, update_memory, delete_memory, and reindex are write/mutation operations with no guidance on error recovery. LLMs do not know: are operations idempotent? Can they retry? Should they confirm? What errors are retryable vs fatal?
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | D | 54 | 2026-07-28+ | v2 |
Search memories using semantic and keyword search with optional filters
Update an existing memory by ID with new title, body, or metadata
No pagination or result-limiting guidance. list_memories and search_memories lack limit, offset, or pagination parameters in descriptions. If vault contains thousands of memories, these tools could return unbounded results, exhausting context window. Pattern requires pagination support.
Descriptions lack WHEN to use context. For example, list_memories vs search_memories distinction is unclear, when should an LLM pick list with filters vs semantic search? Descriptions should explain: list for metadata filtering, search for keyword+semantic matching. Same issue with get_status vs implicit availability.
No guidance on soft-delete reversibility. delete_memory description says 'soft-delete (marks as deleted without permanent removal)' but does not explain: can the agent recover deleted memories? Is there an undo_delete or restore tool? How does deletion affect search results? This ambiguity breaks predictable behavior.
reindex and add_project descriptions are vague. 'Trigger a full reindex' and 'Register a new project' do not explain prerequisites, side effects, or outcomes. How long does reindex take? Does it block other operations? What happens if project path is invalid? These are write operations, they need detailed documentation.