A shared-memory gateway for multiple AI agents. A FastAPI + FastMCP surface in front of a Honcho v3 instance that lets several AI agents read shared context and append changes to one learning memory, so every agent stays in sync.
Server has 5 tools with clear verb-noun naming (recall, remember, log_change, profile_card, recent_changes). All tools have descriptions (10 - 200 chars, within baseline 34 - 392 range). Input schemas are present with type definitions. However, parameter descriptions are sparse or missing for several tools (e.g., 'tags' in remember lacks detail on format/cardinality; 'files', 'reversible', 'backup_path' in log_change are vague). Output schemas are not documented, LLMs cannot predict return structure. Error handling is minimal; no recovery guidance. Security is reasonable (bearer token auth, no secrets in params), but audit logging is basic. Composition is sound (each tool has one responsibility), but chaining IDs are not always returned.
Record a change (Honcho infra-changes session + git ledger).
Compact snapshot of the shared representation for session priming.
Recall what the shared memory knows about a topic (dialectic search).
Return the latest logged changes from the shared ledger.
Store a durable fact/observation in the shared memory.
Output schemas not documented. LLMs cannot predict return structure (e.g., does recall() return plain text or structured JSON with metadata?). Forces agents to guess and parse unstructured responses.
Parameter descriptions are vague or missing. 'tags' in remember() lacks format guidance (array of strings, max length, allowed characters). 'files', 'reversible', 'backup_path' in log_change() are optional but lack constraints or examples.
No error handling guidance. If recall() finds no results, or remember() fails due to storage error, the response does not tell the LLM what to do next (retry, ask user, abort). Agents cannot self-correct.
Inferred effective spec: <=2025-11-25.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | D | 58 | <=2025-11-25 | v2 |
No pagination or result limits documented. recent_changes(n=20) defaults to 20 but does not state max or whether results are truncated. recall() may return unbounded text, risking context window exhaustion.
Chaining IDs not returned. If an agent calls remember() and later needs to reference that memory in log_change(), there is no memory_id returned to link them. Forces extra lookup calls.