Loading ToolBench
Model Context Protocol server that exposes the TrueMemory memory system as tools for Claude and other MCP-compatible AI assistants. Provides agent memory at zero infrastructure cost via SQLite-backed persistent storage with vector search capabilities.
TrueMemory MCP server has 6 tools with complete input schemas and descriptions, but exhibits moderate quality gaps. All tools follow verb_noun naming conventions (recall, store, extract, clear_memories, get_stats, configure_tier), which is positive. Descriptions are present and reasonably detailed (ranging 75-220 chars), meeting the 10-1024 character baseline, though some lack actionable context. Input schemas are properly typed with descriptions for each parameter. However, output schemas are entirely undocumented, the server provides no documentation of what each tool returns, forcing LLMs to infer output structure. Additionally, error handling guidance is absent; tools like clear_memories (destructive operation) lack confirmation/recovery patterns. The recall tool's advanced options (use_reranker, use_hyde, use_agentic) are well-scoped but lack constraints or guidance on when to use each. Overall, the server demonstrates solid foundational quality but lacks production-grade completeness in output documentation and error recovery.
Clear all memories from the persistent store. This is a destructive operation that cannot be undone.
Configure the embedding model tier (base or pro) to control the quality/speed tradeoff for vector search.
Extract and consolidate salient memories from recent conversation context. Uses LLM-guided extraction to identify and compress important information that should be persisted.
Retrieve statistics about the memory store, including total memory count, storage size, and indexing status.
Search and retrieve memories from the persistent memory store based on a user query. Returns relevant memories ranked by relevance, with optional reranking and agentic search capabilities.
Add a new memory to the persistent memory store. Captures a message with metadata including sender, recipient, timestamp, and category.
No output schemas documented. Tools define inputs with JSON Schema but provide zero documentation of return types, fields, or structure. LLMs cannot plan downstream calls or extract the right data without knowing what fields are returned.
clear_memories is destructive but lacks confirmation pattern. The tool accepts a 'confirm' boolean parameter, but no dry-run, undo, or recovery guidance is provided. Agents cannot understand consequences before committing.
No error handling guidance. Tool descriptions do not explain what errors might occur, how to recover, or what the agent should do next. E.g., recall() provides no context about network failures, embedding service downtime, or empty results.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | C | 60 | 2026-07-28+ | v2 |
recall() tool accepts optional parameters (use_reranker, use_hyde, use_agentic) with boolean flags but no guidance on when to use each. Descriptions lack decision heuristics; LLMs must guess which configuration is optimal for different query types.
extract() tool description says 'Uses LLM-guided extraction' but does not specify which LLM or API is required. Dependencies on ANTHROPIC_API_KEY or OPENROUTER_API_KEY are mentioned in module docstring but not in tool description, leaving agents unaware of prerequisites.
get_stats() returns statistics but no indication of what fields or structure comprise the response. 'Storage size' and 'indexing status' are mentioned in description but not formally defined.