Loading ToolBench
Persistent memory for AI agents with FTS5 search, workspaces, auto-capture hooks, and MCP server
14 tools with complete schemas and descriptions, but significant gaps in parameter documentation and error handling. Tool names follow verb_noun convention well (memory_add, memory_search, memory_delete). Descriptions are present but often generic (avg ~80 chars, below 194-char baseline). Most parameters lack descriptions entirely (e.g., memory_list has 'workspace', 'tag', 'limit' with no descriptions). No output schemas documented. No error recovery guidance. No tool annotations (readOnlyHint, destructiveHint). Composition is reasonable, tools are single-purpose, but parameter relationships undocumented.
Add a new memory document
Get smart context with progressive disclosure: pinned + recent + relevant memories, layered by budget
Delete a memory document by ID
Switch active workspace
Retrieve a saved prompt template by name
List memory documents with filters
Save a reusable prompt template by name
Parameter descriptions missing or incomplete. memory_list has 'workspace', 'tag', 'limit' with no descriptions. memory_stats has empty properties. LLMs cannot infer parameter meaning from names alone.
No output schemas documented. Tools return results but LLMs have no visibility into response structure, field names, or types. Breaks downstream tool chaining and forces LLM to guess field names.
No tool annotations (readOnlyHint, destructiveHint, idempotentHint). memory_delete and memory_update are destructive but unmarked. LLMs cannot distinguish safe from risky operations.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | D | 57 | 2026-07-28+ | v2 |
Search memory documents using full-text search
End a session and record summary
Start a new session for tracking agent work
Get memory statistics: document count, workspaces, DB size
Suggest relevant tags for given content based on similar documents
Get documents in chronological order for a date range
Update a memory document's content and/or tags
No error recovery guidance. Tools lack descriptions of failure modes, retryability, or next steps. A failed memory_search returns nothing, LLM has no guidance on whether to retry, refine query, or try different tool.
Parameter constraints not documented. memory_add says 'max 1MB' but no validation rules for tags, workspace names, or content format. memory_timeline requires 'YYYY-MM-DD' format but not enforced in schema.