Long-running agents receive facts with different temporal validity and project relevance. Persisting every interaction can create false writes, stale state, and inconsistent downstream plans.
Use project-aware routing and an explicit selective-update policy to separate current, scheduled, and project state from reusable memory, then replan only when the updated state materially affects action.
The current prototype focuses on personal scheduling and long-horizon planning. I plan to connect this state-and-memory layer to downstream task-execution agents, tools, and domain components, building a broader platform that can act in the world and accumulate long-running interaction traces for more world- and task-grounded benchmarks and evaluation settings.