Loading ToolBench
Persistent memory + context engine for Claude Code via MCP. Spiritual port of memory-lancedb-pro and lossless-claw.
OpenClawdCode has 7 tools with complete input schemas and descriptions, but quality is uneven. Tool names follow verb_noun convention (store_memory, recall_memory, search_vault). Descriptions exist but are brief (avg ~80 chars) and lack actionable context for LLM selection. Parameters have types and defaults, but descriptions are minimal or missing for many (e.g., 'temporal_type' lacks explanation of decay mechanics). No output schemas documented. Error handling is absent, tools return strings with no guidance on failure modes or recovery. store_memory has 11 parameters with complex interdependencies (tier, temporal_type, confidence) that are underdocumented. No tool declares permissions, audit trails, or security boundaries. Composition is reasonable (each tool has one job), but tools lack chaining metadata (e.g., recall_memory doesn't return memory IDs for downstream use).
Auto-extract memories from a conversation summary. Call this at session end with a summary of what was discussed. Extracts up to 5 memories, deduplicates against existing memories, and stores the survivors. Returns a summary of what was stored.
Load a context profile with instructions for a specific use case (e.g. 'repair', 'dev', 'casual').
Write a session summary as a markdown log file.
Search stored memories semantically. Use at session start and when context is needed.
Search a ChromaDB knowledge base for domain-specific information.
Search the indexed Obsidian vault semantically.
No output schemas documented. Tools return strings with no structure. LLMs cannot parse results or chain tools. E.g., recall_memory returns a string but downstream tools need memory IDs.
Parameter descriptions are minimal or absent. 'temporal_type=dynamic makes a memory decay 3x faster' is in tool description, not param description. LLMs cannot infer what 'tier', 'confidence', 'scope' control without explicit param docs.
No error handling or recovery guidance. Tools return strings with no indication of success/failure, invalid input, or what to do next. E.g., search_vault with invalid query returns unstructured text.
Inferred effective spec: <=2025-11-25.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | D | 51 | <=2025-11-25 | v2 |
Store a memory for later recall across sessions and projects. Categories: general, preference, decision, learning, architecture, debugging, tool_usage, profile, preferences, entities, events, cases, patterns. Tier (core|working|peripheral) controls decay floor. Use `core` for durable facts, `working` (default) for active context, `peripheral` for low-signal. temporal_type=dynamic makes a memory decay 3x faster.
Tool descriptions lack WHEN/WHY context. 'Search stored memories semantically' doesn't explain when to call recall_memory vs search_vault vs search_knowledge. LLMs must guess.
No permission gates or audit trails. Tools like store_memory and log_session write state with no access control, logging, or compliance metadata. Violates pattern:permission-gate and pattern:audit-trail.