Agent memory that survives the outage.
BlackBox is agentic-memory infrastructure with two properties most agent memories lack: it survives a full region failure, and it audits what it lets itself remember — every learned fix is gated, deduplicated, and confidence-scored before it can shape recall. We prove both with an incident-response agent working a live outage on CockroachDB.
Agent memory is a distributed-systems problem.
An agent's memory is only trustworthy if it is available during a failure, consistent across regions, compliant with where data may live — and honest about what it lets itself remember. Bolt a vector store onto a cache onto a state store and none of that holds when a region goes dark. BlackBox runs Claude via Amazon Bedrock over one system of record — and everything below reads from the production cluster.
Kill a region. The memory keeps answering.
Every memory replicates across three AWS regions under SURVIVE REGION FAILURE. Lose one mid-incident and recall keeps serving from the survivors — including rows homed in the dead region. The counts above are live from the production cluster.
Institutional memory, in milliseconds
The vector index is distributed like everything else — search runs next to the data in each region, so recall survives whatever the data survives.
Every answer cites its memory
Pinned to its home region
REGIONAL BY ROW domiciles each memory where the incident happened. An EU incident’s data physically stays in the EU — no second database.
Runbooks that compound
Resolving an incident distils what worked into a new runbook, written back to the same memory. The next similar incident recalls it first — the agent gets sharper with every failure it survives.
How the agent works →Reads its own cluster
Through CockroachDB’s Managed MCP Server the agent queries the very database it runs on — mid-drill, it diagnosed its own region outage and reported memory still serving.
Why CockroachDB, in depth →