Loading ToolBench
Intelligent persistent memory for AI agents and IDEs via Model Context Protocol (MCP).
Server has 21 tools with complete input schemas and descriptions. Naming follows verb_noun convention (get_, create_, update_, delete_, find_, search_). Most parameters are typed and described. However, descriptions vary in quality, some are detailed (find_memento_patterns: 280+ chars with usage guidance), others are minimal (get_memento_statistics: 34 chars, 'Handle get_memory_statistics tool call'). Output schemas are not documented in the source. Error handling is present but generic. No tool annotations (readOnlyHint/destructiveHint) despite clear risk classifications. Parameter descriptions lack format/constraint details (e.g., confidence values stated as 0.0-1.0 but no validation guidance). Baseline: avg tool description 194 chars; this server averages ~120 chars with high variance.
Handle adjust_confidence tool call.
Get comprehensive analytics and metrics for the memento graph
Handle apply_confidence_decay tool call.
Handle boost_confidence tool call. Boost confidence when a memory or relationship is used successfully.
Handle create_relationship tool call.
Handle delete_memory tool call.
Find patterns in mementos and relationships Analyzes the relationship graph to detect recurring structural patterns: - Most frequent relationship types - Common memory-type pairs connected by relationships Note: patterns require multiple occurrences of the same relationship type. With small datasets (< 10 relationships), lower the thresholds: min_pattern_size=1, min_support=0.0 → show all types, even singletons. When no patterns meet the thresholds the tool still reports all available relationship types with their counts so the result is always informative.
Output schemas not documented. Tools return results but LLMs cannot plan downstream calls without knowing response structure. Baseline: 100% of A+ tools document return types.
Tool annotations missing. 8 WRITE tools and 1 DESTRUCTIVE tool lack readOnlyHint/destructiveHint. Agents cannot distinguish safe reads from state-modifying calls without explicit hints.
Minimal descriptions on 3 tools. get_memento_statistics (34 chars), get_central_mementos (60 chars), analyze_memento_graph (60 chars) lack context for LLM selection. Baseline: descriptions should be 50-200 chars with WHAT/WHEN/WHY.
Inferred effective spec: <=2025-11-25.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | D | 50 | <=2025-11-25 | v2 |
Find the shortest path between two mementos through relationships
Find mementos that connect different clusters (knowledge bridges)
Handle get_low_confidence_memories tool call.
Handle get_memory tool call.
Detect clusters of densely connected mementos
Get the complete network structure of mementos and relationships
Handle get_memory_statistics tool call.
Handle get_recent_activity tool call.
Handle get_related_memories tool call.
Handle memento_onboarding tool call. Provides comprehensive onboarding protocol for Memento including tool usage guidance, retrieval flow optimization, and best practices.
Handle search_relationships_by_context tool call.
Handle store_memory tool call.
Get intelligent suggestions for relationship types between two mementos
Handle update_memory tool call.
Parameter descriptions lack format/constraint details. Confidence/strength params accept 0.0-1.0 but no validation guidance. Threshold params lack examples. LLMs cannot self-correct invalid input without explicit constraints in descriptions.
No error recovery guidance. Tools return errors but do not guide LLM on next steps (retry, lookup, ask user). Baseline: error responses must state 'what to do next'.