Loading ToolBench
A lifecycle-aware memory layer for AI agents with semantic search, conflict detection, staleness measurement, and temporal reasoning. Provides both a business API (FastAPI) and an MCP HTTP adapter for Agent Memory Leaderboard evaluation.
Minta has 8 tools with mixed quality. Tool names follow verb_noun convention (search, get, append, add) which is good. However, descriptions vary widely in clarity and completeness. The 'add' tool has a detailed schema with proper types and constraints, but several tools lack parameter descriptions or have vague descriptions. Most critically, output schemas are not documented, LLMs cannot predict what fields to expect from responses. Error handling guidance is absent across all tools. The server exposes internal concepts (slot_id, pack, inbox) without explaining their purpose to agents.
Synchronous ingest of messages into memory. Returns 200 only after every message is durable and searchable. Request ID is idempotent.
Liveness check endpoint. Returns health status.
Append a short entry to the inbox after a substantive turn. Never writes into master memory directly — the user confirms from the inbox.
Log a memory read decision before answering about prior work. Ensures the loop read decision is logged for lifecycle tracking.
Load a complete memory pack for a named slot (e.g., a project or active experiment). Returns the whole pack of related context objects.
Search memory for context relevant to a query. Returns ranked, user-scoped evidence for a query (never generates an answer).
Output schemas not documented. LLMs cannot predict response structure (fields, types, pagination). Forces agents to guess what data is available after each call.
Error handling guidance absent. Tools do not explain what to do on failure (retry? ask user? fatal?). E.g., 'search' may return empty results, should agent try a different query or give up?
Domain-specific terminology unexplained. 'slot_id', 'pack', 'inbox', 'autopilot_preflight' are internal concepts. Descriptions do not explain when/why agents should use these tools or how they relate to the memory lifecycle.
Inferred effective spec: <=2025-11-25.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | C | 60 | <=2025-11-25 | v2 |
Simple ping endpoint for connectivity check.
Search memory and return ranked, user-scoped evidence for a query. Never generates an answer, only retrieves relevant context.
'ping' tool has no input schema visible and minimal description (12 chars). Tool appears to be a deprecated liveness check (removed from MCP spec 2026-07-28).
Parameter descriptions incomplete. 'options' in search/minta_search_context is described as 'Optional filter options' but does not explain what filters are valid or how to structure them.