Restart-safe by design: a cache that survives a crash without losing a thing
A cache that forgets everything on restart isn't much of a memory. Crowkis writes durably first, so a crash, a deploy, or a reboot never costs you a single learned answer.
There's a quiet assumption baked into a lot of caches: it's fine to lose everything on restart, because it's 'just a cache.' For an LLM cache that assumption is expensive. Every lost entry is an answer you'll pay the model to regenerate, and every deploy becomes a cold-start tax on your bill and your latency.
In plain words. Crowkis writes to a crash-safe log before it acknowledges a write, then keeps that data in sorted files on disk. Pull the plug and it replays the log on the way back up, nothing learned is lost.
flowchart TD W["CSET"] --> WAL["write-ahead log<br/>append · CRC per record"] WAL --> MT["in-memory table"] MT -- "flush" --> SST["sorted files on disk"] W --> HNSW["vector index persists too"] style WAL fill:#fbe9e8,stroke:#d62221,stroke-width:2.5px
This is what lets Crowkis double as durable agent memory, not just a volatile cache. Years of an agent's learned context can live on disk and come back warm after a restart, because durability was the starting assumption, not a feature bolted on later.
The bottom line
A restart should be a non-event, not a bill. Write durably first, and your cache, and your agents' memory, outlives every crash and every deploy.