Durable State for AI Agents

State that survives a process

Agents need continuity across requests, restarts, and workers. I built shared persistence for sessions and tasks, along with client integrations that let applications use it through their existing frameworks.

The difficult part was agreeing on behavior. Ownership, retries, pagination, and error handling needed to mean the same thing across the service and its clients.

Clear contracts across frameworks

I used shared API contracts and framework-specific adapters. Each integration has an explicit supported surface rather than promising to cover everything a framework can do.

Extending memory carefully

The service now includes long-term memory capabilities in the codebase. My ongoing work explores how to review extracted information, apply corrections, and evaluate retrieval. Newer review tooling is still in development.

I want applications to control what becomes durable memory and to have a clear way to correct it. That requires more than storing conversation history.