Open-source local MCP Server providing cross-application long-term memory for AI applications
OpenMemory has 8 tools with basic schemas and descriptions, but significant gaps limit production readiness. All tools have descriptions (10-60 chars, mostly short), and input schemas are present with type definitions. However, descriptions lack depth, most are single-clause statements without context on WHEN to use the tool or what happens on error. Parameter descriptions are minimal (e.g., 'The UUID of the memory to retrieve' for get_memory). Output schemas are not documented, callers must infer structure from code. Error handling returns JSON with 'error' fields but provides no recovery guidance. No tool annotations (readOnlyHint, destructiveHint, idempotentHint) despite clear risk levels. Naming is verb-first and clear (add_, search_, get_, list_, update_, delete_, get_related, memory_stats), following conventions well. Parameters lack enums where appropriate (e.g., 'type' in add_memory accepts 'fact|preference|episode|entity' but is a free-form string). Defaults are reasonable but update_memory uses importance=-1 as a sentinel, which is fragile. Overall: functional but below production bar for agent reliability.
Store a new memory. Types: fact, preference, episode, entity.
Delete a memory.
Retrieve a specific memory by ID.
Find memories related to a given memory.
List memories with optional filtering.
Get memory statistics.
Search memories by semantic similarity.
Output schemas not documented. Callers cannot infer what fields to expect from responses (e.g., does get_memory return 'id', 'memory_id', 'uuid'?). LLMs must guess structure, risking failed downstream tool chains.
Descriptions are too short (10-60 chars) and lack context. 'Store a new memory' does not explain WHEN to call add_memory vs update_memory, what happens on duplicate content, or what metadata/tags should contain. LLMs cannot reliably select the right tool.
No tool annotations (readOnlyHint, destructiveHint, idempotentHint). delete_memory is destructive but unmarked; search_memory is read-only but unmarked. Agents cannot infer safety properties without explicit hints.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | D | 51 | 2026-07-28+ | v2 |
Update an existing memory.
'type' parameter in add_memory is free-form string but should be an enum (fact|preference|episode|entity). LLMs will hallucinate invalid types; enum constraint prevents this.
Error responses lack recovery guidance. 'Memory not found' tells the LLM nothing about what to do next (search? list? create?). No error classification (retryable vs user-fixable vs fatal).