Loading ToolBench
Brain-inspired memory engine for AI coding agents — three-layer memory, temporal reasoning, hybrid search, consolidation pipelines
Neurox demonstrates strong schema completeness and detailed parameter descriptions across 15 tools. All tools have explicit descriptions (avg ~180 chars) and comprehensive input schemas with enums and type constraints. Naming follows verb_noun patterns (save, recall, update, forget). However, output schemas are not documented in the source, LLMs cannot predict return structures. Error handling guidance is absent; tools do not indicate what to do on failure. Some parameter descriptions are verbose but lack actionable constraints (e.g., 'confidence' lacks min/max bounds). Tool composition is sound (each does one thing), but no evidence of idempotency guarantees or recovery patterns.
Create a safe, consistent backup of the Neurox database using SQLite's online backup API. Works while the server is running. Returns the backup file path and size.
Force immediate memory consolidation: decay, promote, dedup, detect contradictions, reflect, and evict. Normally runs every 30 minutes automatically.
Get relevant context: recent + high-importance + file-linked observations. Use at session start to get proactive memory.
Deep memory curation: review a namespace with a large model to delete noise, recalibrate importance weights, and improve memory quality. Uses the curator LLM provider (e.g. Gemini Flash).
Soft-delete an observation. It won't appear in recall but remains in the database.
Report changed files from a git commit. Marks linked observations as stale.
Output schemas not documented. LLMs cannot predict return structures for any tool. No documented fields, types, or pagination info for responses.
No error handling guidance. Tools do not indicate error categories (retryable, user-fixable, fatal) or recovery steps. Agents cannot self-correct on failure.
Numeric parameters lack bounds. 'confidence' (0.0-1.0), 'limit' (default 10, max 50), 'days' (default 7) have no min/max constraints in schema. LLMs may pass invalid values.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | C | 68 | 2026-07-28+ | v2 |
Compute brain power score (0-100%) showing how much of Neurox's potential is being used. Returns per-dimension breakdown with status and actionable recommendations.
Report an observation as incorrect. Marks it stale, halves confidence. Optionally provide replacement content which creates a new observation with a 'supersedes' link.
Search memories using FTS5 keyword search with tri-factor scoring (recency x importance x relevance). Returns ranked results. Automatically detects temporal intent in queries (current, history, when, duration) and adjusts ranking accordingly. Temporal keywords are stripped from FTS to improve recall.
Trigger reflection: synthesize related observations into high-level insights. Requires LLM provider.
Save an observation to memory. Inserts into Buffer layer with FTS5 indexing. Supports topic_key upsert (same topic_key = update in place).
End a work session with a summary. Marks session as completed.
Start a new work session. Auto-closes previous active sessions. Returns session ID and relevant context.
Get brain statistics: layer counts, staleness stats, and health information.
Update an existing observation by ID.
Destructive operations lack confirmation pattern. 'forget', 'invalidate', 'consolidate', 'curate' modify or delete state without dry-run or confirmation support.
No idempotency guarantees documented. Tools like 'save' with 'topic_key' upsert may be idempotent, but this is not explicitly stated. Agents cannot safely retry.