Loading ToolBench
Multi-agent memory intelligence layer — shared knowledge, expertise tracking, and conflict detection for AI agent teams
Neurohive has 17 tools with explicit schemas and descriptions. Most tools follow verb_noun naming (store_memory, search_memory, recall_memory). Descriptions are present and reasonably detailed (avg ~120 chars), exceeding the 10-char minimum. However, several critical gaps reduce the score: (1) output schemas are not documented, tools return textResult() with generic JSON, but LLMs cannot infer field structure; (2) error handling is minimal, no recovery guidance, no categorization of retryable vs fatal errors; (3) some parameters lack constraints (e.g., similarity_threshold in consolidate has no min/max despite being 0-1); (4) no tool annotations (readOnlyHint, destructiveHint) despite clear risk levels; (5) composition issues, store_memory and search_memory are well-designed, but forget_memory lacks a confirmation pattern for destructive operations. Per-tool analysis shows strong naming and description consistency, but schema completeness and error guidance are weak.
Recompute embeddings for all memories missing from the vector store. Also syncs FTS index.
Run consolidation: merge near-duplicates, decay stale memories, prune low-value, sweep expired, purge old tombstones. Set commit=true to apply.
Create or update an entity in the knowledge graph.
Create or update a relation between two entities in the knowledge graph.
Export memories as JSONL or JSON for backup or migration.
Tombstone (soft-delete) memories matching the given filters. At least one filter is required.
Get claims associated with a specific memory.
Output schemas not documented. Tools return generic textResult() with JSON content, but LLMs cannot infer field structure (e.g., what fields does hivemind_status return? What is the structure of search_memory results?). This forces LLMs to guess and wastes tokens parsing unstructured output.
No tool annotations (readOnlyHint, destructiveHint, idempotentHint). Tools like forget_memory (DESTRUCTIVE), consolidate (WRITE), and import_memories (WRITE) lack explicit risk markers. LLMs cannot distinguish safe reads from dangerous writes without parsing descriptions.
Destructive operations lack confirmation/dry-run pattern. forget_memory and consolidate support dry_run/commit, but hivemind_conflicts resolve_id has no confirmation step. An LLM could accidentally resolve conflicts without preview.
Inferred effective spec: <=2025-11-25.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | D | 52 | <=2025-11-25 | v2 |
List unresolved conflicts between agent memories, or resolve a specific conflict by ID.
Query agent expertise. Provide a topic to find top experts, an agent_id to see their profile, or neither for all expertise.
Get the neurohive dashboard: total memories, agents, categories, recent activity, and conflict count.
Import memories from JSONL or JSON data. Deduplicates by content hash.
Get statistics about stored memories: counts, categories, trust levels, importance, and database size.
Query the knowledge graph for entities and relations.
Recall memories by ID, namespace, category, or tags without semantic search.
Search for claims in the cognitive layer.
Search memories using hybrid vector + full-text search with RRF ranking, graph boost, cognitive priming, and neurohive annotations.
Store a new memory with semantic deduplication, contradiction detection, surprise scoring, entity extraction, and neurohive expertise tracking.
Numeric parameters lack explicit min/max constraints in descriptions. consolidate.similarity_threshold and consolidate.decay_lambda are 0-1 range but descriptions do not state bounds. backfill_embeddings has empty input schema, no parameters documented at all.
Error handling is minimal. No recovery guidance (e.g., 'If memory not found, try search_memory()'), no error categorization (retryable vs fatal), no actionable error messages. LLMs cannot self-correct on failures.