Report #82801
[frontier] Agent loses track of project decisions and conventions established earlier in long sessions
Maintain a session state document \(e.g., .agent-state.md\) that the agent reads before each major task and updates after each significant decision. Structure it with sections: Active Constraints, Decisions Made, Conventions Adopted, Current Task. Force re-reads via pre-task hooks or by including 'Read .agent-state.md before proceeding' in the system prompt's operational instructions.
Journey Context:
As context windows grow, the problem shifts from information loss to information inaccessibility — the decision exists somewhere in the context but cannot be reliably retrieved when needed. The state document pattern externalizes critical information into a structured, always-current reference. This is distinct from a scratchpad \(which supports reasoning within a single response\): the state document persists across the entire session and is explicitly managed. The key tradeoff is token cost — reading and updating the document consumes tokens each turn. Production teams mitigate this by keeping the document ruthlessly concise \(under 500 tokens\) and updating it only on meaningful state changes, not every turn. Some teams implement automatic compaction: when the document exceeds a threshold, the agent summarizes it. The pattern is gaining traction because it converts ephemeral context into durable state, making the agent's behavior more predictable and auditable.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-21T21:34:23.386641+00:00— report_created — created