Cinch is in public alpha.
AI Agent Memory
When agents share a database, one agent's context leaks into another's. That looks like a bug, but it surfaces as a hallucination. Cinch gives every agent its own isolated database that costs nothing when idle.
Agents need state
Conversation history. Tool call results. Knowledge graphs. Session caches. Your agent framework handles orchestration, but the data has to live somewhere between runs.
Shared databases don't work here. One agent's context leaking into another is a bug you don't want to debug at 2am. Key prefixes aren't real isolation — they're a convention that breaks the first time someone forgets.
And the economics are brutal. 10,000 agent sessions on Redis Cloud at $7/GB each is $70K/mo. Most of those agents ran once last Tuesday and haven't been touched since. You're paying active prices for sleeping data.
One database per agent
Spin up a database when an agent starts. Use standard protocols — RESP for cache, Bolt for graph, HTTP for SQL. When the agent goes idle, the database archives itself. When the agent comes back, it wakes in milliseconds.
- → Cache (Redis-compatible) for session state, tool results, conversation history.
- → Graph (Cypher-compatible) for knowledge graphs, entity relationships, RAG context.
- → SQL (SQLite-compatible) for structured logs, metadata, agent configuration.
Active agents pay $0.50/GB on NVMe. Idle agents archive to cloud storage at $0.20/GB. 10,000 agent sessions where 95% are idle: ~$200/mo instead of $70K.
Fits your agent lifecycle
API call returns a connection string. Under a second.
Standard Redis, Bolt, or SQL protocols. Any client library works.
Auto-archives after idle timeout. Cloud storage prices.
First request brings it back. Milliseconds from NVMe.
Clone an agent's state to test different strategies.
Remove the database. No orphaned rows anywhere.