Loading ToolBench
Problem-Solving Memory service — user-owned memory of past problem-solving experience, reusable across AIs and projects. Claude Code plugin runtime for the Memory. Carries the host and protocol knowledge the deterministic adapter deliberately does not have.
Server has 9 tools with explicit schemas and descriptions visible in pre-tool-use.js. Naming follows verb_noun pattern (current_problem, start_problem, add_event). Descriptions are present but brief (10-50 chars for most tools). Input schemas are well-structured with enums, maxLength constraints, and nested objects. However, output schemas are not documented, LLMs cannot predict return structure. Error handling guidance is absent. No tool annotations (readOnlyHint, destructiveHint). Parameter descriptions are minimal; many lack context on when/why to use them. The recall_similar_experience tool has complex nested schema but sparse guidance on structural_features semantics.
Add an event (hypothesis, attempt, dead end, discovery, fix, or user correction) to the current problem
Add a verification result to the current problem
Close the current problem with a final status
Continue working on a previously chosen problem
Get the current problem being worked on in the session, or candidates if ambiguous
Mark the current problem as having a fix candidate
No output schemas documented. LLMs cannot predict return structure or plan downstream tool calls. Tools like current_problem, recall_similar_experience, and add_event lack documented response fields.
Tool descriptions are too brief (10-50 chars). Baseline is 194 chars. Descriptions lack WHEN to use, WHAT it returns, and dependencies. E.g., 'Continue working on a previously chosen problem' does not explain what state changes or what the LLM should expect.
No tool annotations (readOnlyHint, destructiveHint, idempotentHint). Risk field is present in metadata but not exposed to LLM. Agents cannot distinguish safe reads from irreversible writes (close_problem is IRREVERSIBLE but unmarked).
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | D | 55 | 2026-07-28+ | v2 |
Search memory for similar past problem-solving experiences
Resume a paused problem
Start a fresh problem in the current project
recall_similar_experience has complex nested schema (structural_features with 6 array fields) but parameter descriptions are minimal. LLMs cannot infer what problem_domain, symptom_patterns, or suspected_boundaries mean without examples or guidance.
No error handling guidance. Tools lack recovery hints (e.g., 'If problem not found, call current_problem first'). No categorization of errors as retryable vs user-fixable. Agents cannot self-correct on failure.