Durable incident memory for cold-started agents
Most agent memory stores chat history. The thing worth storing under pressure is which remediation step is executing right now — because re-running a remediation action can be worse than never running it at all.
Independent system · architecture, implementation, deployment, and evidence
Continuum runs as a public incident console on Hugging Face Spaces over an AWS Lambda orchestrator and CockroachDB Cloud, deployed from CI on a version tag. The kill-and-resume sequence is reproducible locally and recorded as one unbroken take.
The conditions that cause production incidents — resource exhaustion, node failure, rollbacks, autoscaling churn — are the same conditions that kill the agent responding to them. An agent holding its working state in process memory does not degrade when that happens: it stops, and a human restarts the incident from zero without knowing which remediation actions already ran.
The orchestrator's first action on every invocation — new environment or reused — is a CockroachDB read for open incident state matching the alert. Provisioned concurrency is deliberately absent, so the guarantee never rests on a warm container.
Incident state and a VECTOR(1024) C-SPANN index live in the same CockroachDB cluster, so correlation filters on structured columns and ranks by distance in a single round trip. There is no second database to drift out of sync.
Only the memory agent writes. Each step commits in two explicit SERIALIZABLE transactions — `executing` before the execution window, `executed` after — with the forward step claimed once. A kill lands with `executing` durable, which is the fingerprint the next invocation resumes from.
All incident and alert data is synthetic — this demonstrates a recovery guarantee, not a production incident-response tool. Bedrock correlation and reasoning are best-effort and sit off the recovery critical path, so the flow degrades to deterministic fallbacks rather than failing.
I’m always interested in the constraints that make an architecture worth explaining.