chDB Durable Layer for agent memory

On Sep 28 2026 the ClickHouse team released a durability extension for the embedded chDB engine. The new layer writes a checkpoint of the local analytical database to an object‑storage bucket, allowing the state that an agent accumulates to survive the host that created it. The design keeps the hot‑path queries inside the process while providing a recoverable copy that can be reloaded on any machine. Agent memory that lives only in a MergeTree directory on a single disk cannot be shared across continuous‑integration runs, temporary sandboxes or a second laptop. By persisting the chDB state to object storage the solution bridges the gap between a purely local store and a full‑blown remote ClickHouse server, avoiding network latency, credential handling and the need for a dedicated service. This enables analytics‑driven agents to retain rich, columnar history and still start up quickly on new hosts. The blog notes that the durability layer is a “small durability layer” and that it solves the single‑disk problem, but it does not prove that performance will match a purely in‑process store in all workloads. It remains uncertain whether the approach will scale to multi‑writer scenarios or how it compares to alternatives such as Litestream replication or server‑side ClickHouse deployments. The team’s claims about “full columnar OLAP, recoverable state” are presented without independent benchmarks.