Loading ToolBench
Persistent semantic memory for Claude Code sessions. Stores and searches across all past Claude Code sessions using vector embeddings and PostgreSQL.
Anamnesis has well-structured tool definitions with complete schemas and descriptions. All 5 tools have input schemas with typed properties and descriptions. Tool names follow verb_noun convention (search, recent, session, ingest, report). However, descriptions lack actionable guidance on WHEN to use each tool vs. alternatives, and parameter descriptions are generic. Output schemas are not documented, LLMs cannot predict response structure. Error handling is absent from descriptions. The server targets a specialized use case (Claude Code session memory) which limits composition patterns.
Generate a daily activity report from Anamnesis session data. Returns a markdown report summarizing sessions, topics, and time allocation for a given date. If project is specified, returns a per-project report; otherwise returns a cross-project summary. Requires the "reporting" section in config.
Trigger ingestion of transcript files. Without arguments, discovers and ingests all new/changed files. With session_id, ingests that specific session.
Browse recent Claude Code sessions. Shows session summaries with project, date, tools used, and files touched. Useful for: session continuity, reviewing recent work, finding what was done on a project.
Search YOUR memory — all past Claude Code sessions across all projects. This is how you recall past decisions, discussions, implementations, and debugging sessions. USE THIS PROACTIVELY when the user references past work, asks what was decided, or asks about anything that may have come up in a prior session. Do not say "I don't have context from past sessions" without searching here first.
Get full details of a specific session by ID (partial ID OK). Includes all turns and related sessions. Use when you found a relevant session via search/recent and want the full context.
Output schemas not documented. LLMs cannot predict response structure (fields, types, pagination). Descriptions state WHAT tools do but not WHAT they return.
Parameter descriptions lack format/constraint guidance. E.g., 'since' accepts ISO 8601 but description doesn't specify format. 'limit' has no min/max bounds. 'budget' lacks explanation of token calculation.
No error handling guidance in tool descriptions. LLMs don't know what to do if search returns no results, session_id is invalid, or ingest fails. No recovery hints provided.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | D | 59 | 2026-07-28+ | v2 |
anamnesis_ingest description does not clearly state it modifies state (writes to database). LLMs need to know which calls are idempotent vs. have side effects.
Tool descriptions lack WHEN-to-use guidance. E.g., anamnesis_search vs. anamnesis_recent both retrieve sessions, when should LLM pick one over the other? No disambiguation provided.