Loading ToolBench
Personal Memory Intelligence System with hyperbolic hierarchy, outcome-driven weight evolution, and MCP integration for Claude Desktop and Claude Web
PMIS V2 exhibits significant definition quality gaps. While 19 tools are registered with schemas and descriptions, most descriptions are vague or domain-specific jargon without actionable guidance for LLM selection. Parameter descriptions are sparse or missing entirely. No output schemas are documented. Error handling is minimal, tools return raw dict responses without recovery guidance. Tool naming is inconsistent (memory_turn vs replay_status vs get_daily_summary). The memory hierarchy (Super Context > Context > Anchor) is explained in prose but not formalized in schemas. No tool annotations (readOnlyHint/destructiveHint) despite clear WRITE vs READ_ONLY risk labels in metadata. Composition is poor: memory_turn and memory_store operate on the same domain but lack clear chaining guidance. Productivity tracker tools (add_deliverable, complete_deliverable) lack validation rules and error cases.
Add a new deliverable. Contexts should be comma-separated. Example: add_deliverable("GTM Strategy", "Sales", "Outreach, Collateral", "Rohit", "2026-05-01")
Mark a deliverable as complete. Propagates contribution tracking.
Show what work contributed to deliverables vs what didn't. Includes wasted effort analysis and recommendations.
Get time breakdown for a day: SC → Context → Anchor with human/agent split. Defaults to today. Format: YYYY-MM-DD.
Get progress on deliverables with time + human/agent split. Optionally filter by deliverable name.
Get current tracking status — active segment, window, running time.
No output schemas documented. Tools return dicts with undocumented fields (e.g., replay_summary returns 'convergence_state', 'blend_weight', 'schema_prediction', none described). LLMs cannot plan downstream calls or extract required data.
Parameter descriptions are sparse or missing. 'contexts' in memory_store is described as 'List of context objects with title, weight, and anchors' but no schema for the nested object. 'target_date' in get_daily_summary lacks format specification (YYYY-MM-DD is mentioned in description, not schema). LLMs cannot validate inputs.
No error handling guidance. Tools return {'error': 'No active session for X'} but do not tell LLM what to do next (retry? call another tool? ask user?). No categorization of retryable vs fatal errors.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | F | 49 | 2026-07-28+ | v2 |
Get productivity trends: daily totals, human/agent ratios, top contexts, agent leverage over time.
Browse all super contexts and their hierarchy to see what's stored in memory.
Get memory system statistics (node counts, trees, orphans).
Store a new memory with structured hierarchy: Super Context > Context > Anchor. Use when the conversation produces a new insight worth remembering.
Process a conversation turn and retrieve relevant memories. Call this on EVERY user message to inject memory context into your response.
Pause the productivity tracker. No screenshots will be taken.
Close a conversation's replay session and generate its final summary with feedback aggregation.
Run the full diagnostic report on all collected replay data. Analyzes: short conversation problem, feedback calibration, reranking impact, frustration false positives, and hyperparameter sensitivity. Call after a full day of data.
View the one-bit feedback signal analysis for a conversation. Shows confirmation rate, calibration accuracy, and whether the session engine's schema predictions match user behavior.
Check the replay harness status: uptime, total turns observed, active conversations, collection stats.
Get the session tree analysis for a specific conversation or all active conversations. Shows convergence state, blend weight, schema prediction, and feedback signals.
Resume the productivity tracker after pause.
Search productivity data semantically. Returns matching work entries.
Tool naming inconsistency. Memory tools use verb_noun (memory_turn, memory_store, memory_browse). Replay tools use verb_noun (replay_status, replay_summary). Productivity tools mix get_ and verb_noun (get_status, pause_tracking, add_deliverable). LLMs struggle to infer intent from inconsistent naming patterns.
No tool annotations (readOnlyHint/destructiveHint/idempotentHint). Metadata labels 'READ_ONLY' and 'WRITE' exist but are not exposed in MCP schema. LLMs cannot determine which tools are safe to retry or have side effects.