Local memory server for AI agents using SQLite and MCP
The server implements 25 tools with generally adequate schemas and descriptions, but has significant gaps in consistency, parameter clarity, and error handling guidance. Tool naming follows verb_noun conventions well (remember_fact, recall, create_entity, etc.). Most tools have descriptions between 50-150 characters, meeting baseline expectations. However, parameter descriptions are often minimal (10-30 chars), and output schemas are not explicitly documented in the code. Error handling lacks recovery guidance, tools do not indicate what the LLM should do when a call fails. The tool set is large but composition suffers from overlapping concerns (remember_fact and remember_facts do similar things; add_todo and add_task are near-duplicates). Critical issue: destructive operations (forget, delete_entity, delete_relation, delete_task, delete_observation) lack confirmation or dry-run support, violating the confirmation-request pattern. No tool annotations (readOnlyHint, destructiveHint) are visible in the schema definitions.
Add a task to a conversation or global task list
Add a new todo item
Cluster memories based on semantic similarity
Mark a todo as complete
Consolidate and summarize memories into a context summary
Create a new entity in the knowledge graph
Create a relationship between two entities
Delete an entity from the knowledge graph
Delete an observation from an entity
Destructive operations lack confirmation or dry-run support. Tools like forget, delete_entity, delete_relation, delete_task, and delete_observation can permanently destroy data without a confirmation step or preview.
Tool annotations missing. No readOnlyHint, destructiveHint, or idempotentHint tags visible in schema definitions. This prevents MCP clients from recognizing which tools are safe vs destructive.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-21 | D | 50 | 2026-07-28+ | v2 |
| 2026-03-09 | F | 47 | - | v1 |
Delete a relationship between two entities
Delete a task
Export all memories to a JSON file
Delete a memory from the database
Initialize a new conversation with optional name
Retrieve recently created or accessed memories
List tasks, optionally filtered by conversation or status
List all todos, optionally filtered by status
Query the knowledge graph of entities and relationships
Query memories using semantic search or keyword matching
Provide explicit reinforcement signal to a memory to increase its importance
Store a fact or observation in memory with optional tags
Store multiple facts or observations in memory at once
Maintains memory lifecycle importance using explicit reinforcement, bounded decay, and immune tags. Passive recall has no effect.
Update an existing entity in the knowledge graph
Update the status of a task
Overlapping tool functionality. remember_fact and remember_facts perform similar operations (store facts) with minor differences. Similarly, add_todo and add_task are near-duplicates. This forces LLMs to reason about which to use, violating single-responsibility and composition principles.
Error responses lack recovery guidance. No evidence in tool definitions that error messages explain what the LLM should do next (e.g., 'Entity not found. Try search_entities() first'). Errors appear to be raw or unstructured.
Output schemas not documented. Tools like recall, list_recent_memories, and cluster_memories do not explicitly declare what fields they return. This forces LLMs to guess the response structure and breaks downstream tool chaining.
Parameter descriptions are minimal (10-30 chars). Examples: 'Maximum number of results to return' (35 chars, acceptable), but 'Include outdated memories in results' lacks context on WHEN to set it true vs false. LLMs cannot infer intent from bare boolean flags.
soul_maintenance tool accepts empty input {}. Its purpose ('Maintains memory lifecycle importance...') is explained, but it is unclear when to call it, how often, or what state it modifies. Appears to be an internal maintenance operation, not a user-facing action.
Inconsistent parameter naming. Some tools use 'limit', others use 'num_clusters'. Some return 'memory_id', others might return 'id'. This forces the LLM to map field names across tools, increasing error likelihood.