AML integration contract

This document records the public Agent Memory Leaderboard contract checked on 2026-08-04. Recheck the official documentation before a formal run because the organizers may update operational limits.

Official sources:

• https://agentmemories.ai/api-guide
• https://agentmemories.ai/docs
• https://agentmemories.ai/rules

Endpoints

• GET /health is unauthenticated and returns any HTTP 2xx when ready.
• POST /add is synchronous. HTTP 200 means every submitted message is persisted and immediately searchable.
• POST /search returns {"data": [...]} in relevance order and never returns more than the requested top_k (100 in formal external evaluations).

user_id is the sole retrieval-isolation boundary. session_id identifies a source conversation but is not a Search filter. Add responses must echo request_id, user_id, and session_id exactly.

The hosted API accepts Authorization: Bearer, Authorization: Token, or X-Api-Key. This implementation supports all three. Keep the configured key out of the repository and logs.

Data handling

Evaluation payloads and derived data may only be used for the current run. They must not be used for training, analytics, reconstruction, or sharing and must be deleted within 30 days after the run unless the organizer grants a different retention period in writing.

Local versus official results

The scripts in this repository validate the public HTTP contract and replay normalized public benchmark data. They do not reproduce the private suite, locked Answer model, evaluator configuration, or official aggregate score. Only the organizer's Smoke and Full jobs are official.