MCP server that gives several agents one shared, versioned memory — plus an explicit handoff packet so the next session starts where the last one stopped.
Seven tools with solid naming (verb-first: remember, recall, history, handoff, resume, record_outcome, memory_stats) and comprehensive descriptions (avg 150+ chars). Input schemas present with Zod validation and type constraints. However, output schemas are NOT documented, responses are JSON-stringified without formal schema declarations. Parameter descriptions are good but some lack constraint details (e.g., 'refs' array in handoff/resume lacks element type docs). Error handling is minimal, no recovery guidance or actionable error messages visible in code. Tool annotations present (readOnlyHint) but incomplete (missing destructiveHint on write tools). No pagination support despite potential for large result sets.
Write a handoff packet: what was done, what is left, what is still unclear, and which exact record versions the next agent must read. With no refs given, everything this session touched is pinned.
Every version of one key, oldest first: who wrote it, when, with what confidence, and the hash chain that ties the versions together.
Counts per agent and scope, average confidence, handoff states, and an integrity check of the hash chain.
Read the current version of a key (scope+key), or search current records by scope prefix, tag, free text and minimum confidence.
Close the loop: report success or failure against the records that drove the decision. Their confidence moves accordingly, and the before/after values are kept as an audit trail. Pass handoff_id to close the packet in the same call.
Store a fact under scope+key. Writing an existing key appends a new version instead of overwriting, so nothing is ever silently lost.
Output schemas not documented. Tools return JSON-stringified responses without formal schema declarations. LLMs cannot predict response structure for downstream planning.
Missing destructiveHint annotations on write tools. 'remember', 'handoff', 'resume', 'record_outcome' modify state but lack destructiveHint=true, preventing LLMs from recognizing irreversible operations.
No error handling guidance. Code does not show error responses with recovery hints (e.g., 'Record not found. Try recall() with a broader scope'). Errors will be raw exceptions.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | C | 60 | 2026-07-28+ | v2 |
Claim the oldest open packet addressed to this agent (or a specific one by id) and get back the packet together with the pinned records. Refs whose version has moved on are reported as stale rather than silently swapped for the newer version.
Parameter 'refs' in handoff/resume lacks element type documentation. Array of what? RecordRef objects? Strings? LLMs cannot construct valid payloads without explicit schema.
No pagination support. 'recall' search can return unbounded results (limit capped at 200 but no offset/cursor). Large result sets will bloat context and degrade LLM reasoning.