Loading ToolBench
Local-first AI agent memory system. No account needed. Provides MCP tools for session management, semantic search, memory recording, and structured data queries.
Awareness Local provides 9 tools with mostly complete schemas and descriptions. Naming follows verb_noun convention (awareness_*). Descriptions range 50-250 chars, meeting baseline. However, several tools have vague or incomplete descriptions (e.g., awareness_publish_agent lacks any description), and parameter descriptions are inconsistent. Some parameters lack type constraints (e.g., 'agent_role' is a free-form string with no enum). Error handling guidance is absent, tools do not explain recovery paths. Output schemas are not documented in the tool definitions. The awareness_record tool conflates 5 distinct actions (remember, remember_batch, update_task, submit_insights, write) into one tool, violating single-responsibility principle.
Retrieves and applies a learned skill by ID. Returns skill methods, trigger conditions, linked source cards, and guidance for execution. Updates usage count and decay score.
Spec-based agent prompt. Returns the system prompt and role-specific instructions for a given agent role from the awareness-spec.json.
Session creation + context loading. Initializes a new session and loads relevant context (cards, tasks, perception signals) based on time window and focus query.
Type-based structured data queries. Retrieves specific data types (context, tasks, knowledge cards, risks, session history, timeline, perception signals, skills) with optional filtering.
Records the outcome of a skill execution (success/partial/failed). Updates decay score, confidence, and consecutive failure count. Marks skill as 'needs_review' after 3 consecutive failures.
Publishes an agent configuration or skill set for sharing or deployment.
awareness_publish_agent has no description and no visible input schema. Cannot determine when/why to call it or what parameters it accepts.
awareness_record conflates 5 distinct actions (remember, remember_batch, update_task, submit_insights, write) into one tool via an 'action' enum. Violates single-responsibility principle, should be split into separate tools (create_memory, batch_record_memories, update_task, submit_insights, write_memory).
Parameters like 'agent_role' (awareness_get_agent_prompt, awareness_record) are free-form strings with no enum constraint or validation guidance. LLMs will hallucinate invalid role names. Should define valid roles as an enum or provide discovery tool.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | D | 59 | 2026-07-28+ | v2 |
Progressive disclosure search (summary/full). Performs hybrid semantic and keyword search across memories, knowledge cards, and tasks with optional context expansion.
Remember / remember_batch / update_task / submit_insights. Records memories, batches of items, task updates, or extracted knowledge insights into the local memory system.
Searches workspace graph nodes (code symbols, files, documentation) by query. Returns node metadata with optional neighbor expansion for context.
No tool provides error recovery guidance. Descriptions do not explain what to do if a call fails (e.g., 'If recall returns no results, try awareness_lookup with broader scope'). Agents cannot self-correct.
Output schemas are not documented in tool definitions. Descriptions do not specify what fields are returned or their types. LLMs cannot plan downstream tool calls or extract required data.