Loading ToolBench
Independent dynamic state engine with remote MCP, OAuth, short-term continuity and optional memory adapters. Ombre Brain is a vendored memory system with neural subsystems, tool output contracts, and projection observers.
11 tools with metaphorical names (breath, hold, grow, trace, anchor, pulse, plan, dream, letter, I, forget) that obscure their actual function. Most lack clear verb-noun naming (e.g., 'breath' should be 'retrieve_memory' or 'surface_memory'). Descriptions are present but vague and poetic rather than actionable for LLM tool selection. Input schemas visible with basic types and descriptions, but lack constraints (enums, ranges, patterns). No documented output schemas. Error handling and recovery guidance absent. Tool composition unclear, relationships between memory operations not specified.
Self-description and identity memory operations
Create stable reference points in memory (formerly release)
Surface memory and emotional state from buckets
Generate and process dream-like memory consolidations
Manage memory erasure and tombstone records
Update and evolve existing memories
Store and retain memory in permanent buckets
Tool names are metaphorical and do not follow verb_noun convention. 'breath', 'hold', 'grow', 'trace', 'anchor', 'pulse', 'dream', 'letter', 'I' obscure function. LLMs cannot infer intent from names alone. Should be 'retrieve_memory', 'store_memory', 'update_memory', 'resolve_conflict', 'create_anchor', 'check_health', 'generate_consolidation', 'manage_letter', 'query_identity'.
No output schemas documented. LLMs cannot plan downstream tool calls or extract required fields (e.g., bucket_id for chaining). Responses are unstructured.
Inferred effective spec: <=2025-11-25.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | F | 44 | <=2025-11-25 | v2 |
Read and write letter-based memory communications
Create and manage action plans
Check homeostatic signals and system health
Trace and resolve memory conflicts or tensions
Descriptions are poetic and vague (10-50 chars). 'Surface memory and emotional state from buckets' lacks actionable context. No guidance on WHEN to use each tool vs. similar ones. Descriptions should be 50-200 chars and answer: what does it do, when to use it, what does it return?
Input parameters lack constraints. 'domain' and 'subsystem' are free-form strings with no enum, pattern, or validation guidance. 'steps' in 'plan' is an array with no item schema. Unbounded inputs invite hallucinated values.
No error handling or recovery guidance. Tool definitions do not specify what errors are possible, whether they are retryable, or what the LLM should do next. No actionable error messages documented.