Loading ToolBench
MCP Server adapter for memory-claw - Universal Memory & Context Engine for LLMs. Exposes memory search and retrieval tools via Model Context Protocol with support per-user isolation.
MemoryClaw exposes 5 well-named tools with reasonable descriptions and visible JSON Schema inputs. However, several parameter descriptions are sparse or missing, output schemas are not formally documented, and error handling guidance is absent. The tools follow a clear verb_noun naming convention (memory_search, memory_get, memory_store, memory_delete, memory_context), which is strong. Most parameters have type definitions and basic descriptions. The server targets a specialized domain (LLM memory management) and implements user isolation via userId, which is a good security pattern. Compared to production baselines (90% of A+ tools have detailed param descriptions, 100% have output schemas), this server is competent but incomplete. It lands in the 'Fair' range, definitions are present but lack depth in parameter constraints and output documentation.
Read or write the active working memory (scratchpad) for a user. The context is a JSON object stored per-user at context.json. Use 'get' to read current context, 'set' to shallow-merge new data into it.
Delete a memory file from the workspace. Use this to remove outdated or incorrect information. Cannot delete MEMORY.md (the main memory file).
Retrieve specific lines from a memory file. Use this after memory_search to get full content of relevant sections.
Search long-term conversation memory using hybrid vector + keyword search. Searches across MEMORY.md and memory/*.md files in the workspace.
Save or update a memory file in the workspace. Creates directories automatically if they don't exist. Use this to persist new information, conversation summaries, or learnings.
Output schemas are not documented for any tool. LLMs cannot reliably parse responses or plan downstream calls without knowing what fields to expect.
Parameter descriptions are generic and lack actionable constraints. Users cannot infer valid ranges, formats, or error recovery paths from descriptions alone.
Error handling is missing. Tools have no documented recovery paths, agents cannot determine if an error is retryable, user-fixable, or fatal.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-21 | D | 57 | 2026-07-28+ | v2 |
| 2026-03-09 | F | 47 | - | v1 |
memory_context uses a generic 'data' parameter of type 'object' with no documented schema. LLMs cannot know what fields to include in a 'set' call.
memory_search accepts maxResults and minScore with numeric defaults (6, 0.15) but provides no justification or tuning guidance. LLMs may pass arbitrary values that break search.