Loading ToolBench
Exposes a persistent memory graph as MCP tools over stdio. Stores facts about people, projects, features, meetings, and decisions using Graphiti and FalkorDB, with automatic session summarization and long-term recall.
Two tools with clear verb-noun naming (recall, remember) and reasonable descriptions (120-180 chars). However, critical gaps: parameter descriptions are minimal or absent, no output schemas documented, no error handling guidance, and no input validation constraints. The 'when' parameter in remember lacks format specification despite accepting ISO datetime. Schemas are inferred from type hints rather than explicitly defined. No pagination, no structured output documentation.
Recall facts from Mnemo's memory about people, projects, features, meetings, decisions, or anything stored earlier. Call this BEFORE answering questions about the user's world or history. Returns fact statements with their timing.
Store new information into Mnemo's memory graph. Pass a natural-language statement (who did/said/owns what). `when` is an optional ISO date/datetime the fact refers to (defaults to now). Entities and relationships are extracted automatically and persisted across sessions.
Output schemas not documented. recall() and remember() return strings with no structure specification. LLMs cannot plan downstream operations or extract typed fields.
Parameter 'when' in remember() lacks format specification. Description says 'Optional ISO date/datetime' but does not state accepted formats, timezone handling, or validation rules. LLMs will guess.
No error handling guidance. Tools return plain strings ('Remembered.') with no recovery hints. If recall() finds no facts or remember() fails, LLM has no actionable error message.
Input validation absent. recall() accepts any string; remember() accepts any 'when' value. No constraints on query length, date format, or entity count. LLMs can pass invalid ISO dates or unbounded queries.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | F | 47 | 2026-07-28+ | v2 |
No pagination or result limits documented. recall() may return unbounded fact lists, risking context window exhaustion. No mention of max results or cursor-based pagination.