Agent State and Memory

Agent memory should be an explicit state object that can be serialized, replayed, and inspected. Chat history is not enough because it loses typed evidence, tool provenance, and quality gate status.

The Problem

Long-running research agents make dozens of tool calls and intermediate judgments. If the only memory is the prompt transcript, the run cannot be resumed safely, audited by a reviewer, or compared against an ablation. The agent may also reuse stale claims because evidence has no timestamp or source boundary.

The Pattern

WRONG

messages = []
messages.append({"role": "user", "content": task})