Promise Ledger: Agent Memory That Keeps Promises

Durable, scoped agent memory that keeps customer promises visible until one person owns the next move.

Comment

Inspiration

Small service teams make promises in calls, chat, and handoffs. The message may survive, but responsibility often does not. Promise Ledger treats useful agent memory as accountability—not indiscriminate transcript storage.

What it does

Promise Ledger turns each customer commitment into a durable operational record with its source, due state, accountable owner, customer and project scope, and an ordered event trail.

Its decision policy is intentionally legible:

  • No owner → ASSIGN_OWNER
  • Owner but no eligible customer-plus-project context → HOLD_FOR_CONTEXT
  • Owner plus scoped context → PREPARE_REVIEW_DRAFT

A prepared draft still requires owner review. The public agent never sends a customer message automatically.

How we built it

The React interface opens on the human consequence: a customer is waiting and one person must own the next move.

CockroachDB Cloud stores promise records, ordered events, and customer-plus-project-scoped memory. The project meaningfully uses the CockroachDB Cloud Managed MCP Server for cluster-scoped read-only inspection and the agent-ready ccloud CLI to inspect the real cluster and verify the fixture replay.

AWS Lambda and HTTP API run the read-only proof service in us-east-1. Public requests verify CockroachDB health, read a promise, and return a decision. API Gateway rejects public owner changes because state changes require an authenticated operator route. Responses include request IDs while logs avoid customer content.

GitHub Actions runs lint, the production build, policy tests, schema validation, and fixture validation.

Challenges

  • Making memory useful without unrestricted transcript retention.
  • Keeping the agent helpful while refusing to invent context or ownership.
  • Proving sponsor integrations as judge-visible behavior rather than decorative service names.
  • Exposing live cloud evidence while keeping public writes closed.

Accomplishments

  • A real CockroachDB-backed decision trail with three testable outcomes.
  • Managed MCP and ccloud proof against the same persistent store.
  • A live AWS Lambda/API path for health, records, and decisions.
  • A deliberate public-write boundary that fails closed.
  • Reproducible tests and a 54-second narrated judge film.

What we learned

Agent memory becomes production-worthy when it preserves the evidence behind a decision and makes the next accountable action visible. Keeping more text is not the same as remembering well.

What's next

Add an authenticated operator surface, measure promise-resolution time with a small service team, and test whether the ledger reduces missed commitments without increasing unnecessary customer messages.

Try it

Built With

  • amazon-api-gateway
  • aws-lambda
  • ccloud-cli
  • cockroachdb
  • cockroachdb-cloud
  • model-context-protocol
  • node.js
  • react
  • typescript
Share this project:

Updates