Loading ToolBench
Persistent AI memory for Claude, Cursor, and other MCP clients
OneBrain MCP has 9 well-named tools with clear action verbs (get_, search_, write_, list_, upsert_, deep_search, skill_feedback). All tools have descriptions (avg ~120 chars, within baseline 34-392 range). Input schemas are present and mostly complete with enums and constraints. However, output schemas are not documented in the provided source, critical for LLM planning. Parameter descriptions are present but some lack detail on constraints (e.g., deep_search 'alpha' parameter lacks explanation of fusion semantics). Error handling is mentioned as 'errorReporting=true' but no recovery guidance visible in code. Tool composition is sound (single responsibility, good chaining via IDs), but batch tool (write_memory_batch) lacks per-item error reporting.
Semantic search across memories using DeepRecall. Supports keyword, vector, or hybrid mode. Returns ranked results with relevance scores.
Get project details or list active projects.
Get the full brain context: profile, recent memories, top entities, active projects, and stats.
List learned skills from SkillForge. Skills are patterns the brain has extracted from memory and usage. Requires Pro or Team plan.
Search and filter memory items by type and status.
Record feedback on a skill — whether it was applied, referenced, or dismissed. This improves skill confidence scoring over time.
Output schemas not documented. LLMs cannot plan downstream calls or extract required fields (e.g., what does get_user_context return? Does it include entity IDs for chaining?). Critical for multi-step agent workflows.
skill_feedback 'context' parameter is typed as object with no schema. LLMs cannot determine what fields to include. Should define a structured schema or provide examples of valid context objects.
deep_search 'alpha' parameter (fusion weight) lacks explanation of what 0 vs 1 means semantically. Description should clarify: '0=keyword-only, 1=vector-only, 0.6=balanced hybrid'.
write_memory_batch returns no per-item success/failure. If 1 of 10 items fails, agent cannot retry just that item or understand which succeeded. Should return array of {item, success, error} objects.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | C | 61 | 2026-07-28+ | v2 |
Create or update an entity (person, place, organization, etc.).
Create a new memory item via extraction. Creates a candidate memory and source event.
Create multiple memories at once (1-10). More token-efficient than individual calls. Duplicates auto-detected.
No confirmation/dry-run pattern for destructive writes (write_memory, upsert_entity, skill_feedback). Agents can accidentally overwrite or delete memories without verification.