flowgrid-memory-runtime

Inspiration

AI agents can remember conversations but still lose why a project made a decision. Worse, they may treat an unreviewed inference as fact or revive a rejected direction. FlowGrid Memory Runtime helps agents inherit trusted judgment state, not just longer context.

What it does

The runtime tracks project judgments as confirmed, pending, rejected, or superseded, with source evidence, revision requests, handoffs, and an audit trail. New evidence can propose a revision, but it cannot silently overwrite a confirmed decision.

How we built it

CockroachDB Cloud Basic is the persistent memory layer. It stores projects, source events, judgments, revisions, evidence links, handoffs, audit events, and vector-searchable judgment summaries. CockroachDB Distributed Vector Indexing supports semantic recall, while CockroachDB Managed MCP was used to inspect the live schema and synthetic lifecycle state.

The runtime is deployed on AWS. AWS Lambda executes the controlled lifecycle transaction, API Gateway exposes a public read-only Runtime Snapshot, a private S3 bucket stores synthetic run traces, and CloudWatch retains logs for seven days. The browser never receives the CockroachDB connection URL or the runtime write token. The deployed verifier completed initialization, revision proposal, evidence application, supersession, vector recall, and final-state retrieval against CockroachDB.

Challenges we ran into

The hardest problem was separating raw information, agent inference, and human-authorized decisions. Typical memory systems compress them into a fluent summary. We instead needed new evidence to trigger revision while preserving the prior decision, its rationale, and the exact transition that superseded it.

We also had to keep a public demo inspectable without exposing database credentials or an unrestricted mutation endpoint. That led to a split between the authenticated Lambda write path and the browser's read-only snapshot path.

Accomplishments that we're proud of

We implemented and deployed a complete governed-memory lifecycle using synthetic data: initialization, agent handoff, pending revision, evidence-authorized supersession, vector retrieval, and an auditable history. We verified the CockroachDB vector index, inspected the live state through Managed MCP, deployed the AWS stack, and published both a browser demo and a public video under three minutes.

What we learned

Agent memory is not just about remembering more. The important question is whether an agent knows what it may trust, what still needs review, and what evidence can change an existing judgment. Provenance must begin when raw input enters the system, and every state transition needs an explicit authorization boundary.

What's next for flowgrid-memory-runtime

Next we will measure continuation accuracy across isolated agent sessions, add operational hardening such as rate limits and production-grade secret management, and test the governed-memory protocol with longer synthetic workflows before considering any non-synthetic deployment.

Built With

• agent-memory
• ai-agents
• audit-trail
• aws-lambda
• cockroachdb
• decision-provenance
• mcp
• multi-agent-systems
• vector-search