Resolve Memory Agent

Inspiration

Agents are only as reliable as the evidence they can remember. Resolve Memory Agent turns a small evidence record into a durable, auditable case memory instead of treating each request as a one-off answer.

What it does

A client submits evidence for a case. The agent checks required fields, hashes the evidence, decides whether the case is verified or needs review, and stores both the resolution and the decision event. Repeating the same case demonstrates recall: the stored decision and history become part of the next response.

How we built it

The public demo is an AWS Lambda Function URL backed by CockroachDB. Lambda handles the HTTP interface and the resolution logic; CockroachDB stores structured case state and an append-only event history. The API exposes POST /api/resolve for evaluation, GET /api/cases/{case_id} for recall, and a human-review route for follow-up decisions.

The database adapter uses the PostgreSQL-compatible CockroachDB wire protocol. We packaged the service for Lambda and verified the deployed flow end to end: create a case, persist its evidence hash and event, then retrieve the case and history.

Challenges and learnings

The biggest challenge was making the deployment reproducible in Lambda. Native PostgreSQL client packaging created a runtime-loading problem, so we replaced it with the pure-Python pg8000 adapter and kept the same CockroachDB SQL model. That made the final artifact smaller and removed a fragile native dependency while preserving TLS and transactional writes.

What's next

Next we would add richer evidence schemas, similarity search over prior cases, authentication and rate limits, and a dashboard for human reviewers. The foundation is intentionally small: an agent that can remember why it made a decision and show the evidence behind it.

Built With

• python