Loading ToolBench
Local-first AI agent memory system with MCP server support for cross-session recall, knowledge cards, skills, and task management
Awareness Memory SDK provides 10 tools with mostly complete schemas and descriptions. Naming follows verb_noun convention (awareness_*). Descriptions are present but vary in quality (50-300 chars). All tools have input schemas with types and descriptions. However, output schemas are not documented in the provided source, parameter descriptions lack actionable constraints (ranges, formats, enums), and error handling guidance is absent. Tools like awareness_record combine multiple actions (remember, batch, update_task, submit_insights) which violates single-responsibility. No tool annotations (readOnlyHint/destructiveHint) despite clear read/write distinctions.
Workflow checklist tool. Returns step-by-step guidance for using Awareness memory tools. Low-cost call for reminding users of the correct workflow sequence.
Apply a learned reusable skill by ID. Returns skill details including methods, pitfalls, verification steps, and linked source cards. Increments usage count and evaluates growth stage.
Spec-based agent prompt generation. Returns role-specific system prompt for agent configuration.
Session creation and context loading. Returns session_id, past context cards, active skills, and user preferences for the current session.
Type-based structured data queries. Returns context dump, open tasks, knowledge cards, risk items, session history, timeline events, perception signals, or learned skills.
Record skill usage outcome (success/partial/failed). Updates decay score, confidence, and consecutive failure count. Marks skill as 'needs_review' after 3 consecutive failures.
awareness_record combines 5 distinct actions (remember, remember_batch, update_task, submit_insights, write) in one tool. Violates single-responsibility principle. LLM must reason about which action to invoke rather than calling a focused tool.
Output schemas not documented. Tools return complex nested objects (e.g., awareness_record with knowledge_cards, action_items, risks, skills) but no schema definition visible. LLMs cannot plan downstream calls or extract fields reliably.
Parameter descriptions lack actionable constraints. 'days' has no range (1-365?), 'limit' has no bounds (1-100?), 'recall_mode' enum is documented but other enums lack explanation of when to use each mode.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | C | 61 | 2026-07-28+ | v2 |
Publish agent configuration and learned skills to the market. Enables sharing of agent profiles and reusable procedures.
Progressive disclosure search with hybrid semantic/keyword modes. Returns summary (~50-100 tokens) or full content for specified IDs. Supports multi-level context retrieval and topic-based expansion.
Record memories with automatic action inference. Supports remember (single card), remember_batch (multiple cards), update_task (task status), and submit_insights (pre-extracted knowledge with cards, action items, risks, and skills).
Search workspace graph nodes by query. Returns node summaries with optional neighbor expansion. Supports filtering by node type and limiting results.
No error handling guidance. Tools declare risk levels (READ_ONLY, WRITE) but provide no recovery instructions. E.g., if awareness_recall returns no results, should LLM retry with different query or expand scope?
__awareness_workflow__ has dunder naming convention and minimal schema (empty object properties). Appears to be a utility/meta-tool but lacks clear purpose and description.