Persistent, local-first memory for AI agents. Selective by design: filters what's worth keeping, links what's related, and says nothing when it doesn't know. SQLite + ONNX, no cloud, no API keys.
AWM provides 6 coordination tools with consistent naming (coord_* prefix) and explicit Zod schemas. All tools have descriptions and input parameters are typed. However, descriptions are terse (avg ~80 chars, below the 194-char baseline), output schemas are not documented, and error handling lacks recovery guidance. Parameter descriptions are minimal, many lack context about valid ranges, formats, or dependencies. No tool annotations (readOnlyHint/destructiveHint) despite clear WRITE vs READ_ONLY risk classifications. Missing idempotency guidance for retry-heavy coordination workloads.
Create a task assignment. Orchestrator uses this to dispatch work to agents.
Get your current assignment, or auto-claim the next pending task. Call this to find out what to work on.
Report progress or completion of an assignment. Set status to in_progress, completed, failed, or blocked.
Register or heartbeat an agent in the hive. Call at session start and periodically.
Sign off from the hive. Releases all locks and marks agent as dead. Call at session end.
Acquire a file lock to prevent concurrent edits by other agents.
Output schemas not documented. LLMs cannot predict return structure (e.g., does coord_checkin return agent_id, status, timestamp?). Breaks tool chaining and forces agents to guess field names.
No tool annotations (destructiveHint, readOnlyHint, idempotentHint). Risk classification exists in metadata but is not exposed to LLM. Agents cannot distinguish safe reads from destructive writes without reading descriptions.
Parameter descriptions lack format/constraint details. E.g., 'agent_id' says 'Your agent ID from coord_checkin' but does not state it is a UUID or how to obtain it if missing. 'status' enum values undocumented in coord_assignment_update.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | C | 65 | 2026-07-28+ | v2 |
Error handling not visible in source. No recovery guidance (e.g., 'If agent_id not found, call coord_checkin first'). Agents will retry blindly or give up.
Idempotency not declared. coord_checkin is a heartbeat (safe to retry), but coord_assign and coord_lock may have side effects. No guidance on retry safety or duplicate prevention.