Continuum

• 

  The durable task cursor lives in CockroachDB, not in process RAM.

• 

  Invocation A commits two steps, records the crash boundary, and stops.

• 

  Invocation B reconstructs the same session from CockroachDB and finishes steps 3–4.

• 

  Continuum — the runtime stops. The mission continues.

Inspiration

Production agents do not fail neatly between jobs. They lose compute while tools are running, tasks are half-done, and the next action lives only in process memory. Restarting from zero can duplicate side effects. Asking a human to re-brief the mission defeats autonomy. Continuum treats durable memory as execution correctness, not as chat history.

What it does

Continuum is a crash-resilient agent runtime. The public proof creates a four-step incident-response mission, writes session, tasks, results, events, and semantic memories to CockroachDB, then stops Invocation A after step two. A fresh Invocation B receives only the client-scoped session UUID, reconstructs the mission from CockroachDB, skips finished work, recalls Bedrock-embedded memory, and continues at step three.

How we built it

The task cursor, append-only event trail, and vector memories live in CockroachDB Cloud on AWS. Amazon Bedrock Titan Text Embeddings V2 turns memory text into 1024-d vectors stored beside the workflow state. Distributed Vector Indexing keeps semantic recall in the same transactional database as the cursor. The Cloud Managed MCP Server lets us inspect schema and live rows during recovery without a custom proxy. The live demo is a Workers app at continuum.vortex-digital.de.

Challenges we ran into

Most agent demos keep workflow position in RAM, so a crash demo is fake unless the next request is truly stateless. Continuum had to prove Invocation B starts cold and still continues. Public-demo isolation was also hard: every visitor needs their own session without leaking another visitor's cursor or memories.

Accomplishments that we're proud of

A real crash boundary: Invocation A commits two steps and dies. Invocation B reconstructs from CockroachDB only and finishes steps three and four. Vector memory and task state stay in one database, so there is no dual-write drift after failure.

What we learned

Agent memory is not a log you bolt on later. If the cursor, events, and embeddings are not in the same durable store, recovery becomes guesswork. CockroachDB's vector index next to SQL state made that one system of record.

What's next for Continuum

Turn the four-step proof into a reusable durable-workflow SDK, add idempotency keys for real tool calls, and exercise regional failover on a multi-region CockroachDB cluster.

Built With

• amazon
• bedrock
• cloudflare
• cockroachdb
• javascript
• mcp
• node.js
• postgresql
• vector
• web