Report #62300
[architecture] Agent loses track of current task goals during long context windows
Maintain a structured scratchpad or working memory block at the very top or bottom of the prompt, explicitly tracking current goals, available tools, and task progress, separate from retrieved long-term memories.
Journey Context:
As context length increases, LLMs suffer from 'lost in the middle' syndrome. If the overarching goal is buried in long-term memory or the middle of the prompt, the agent drifts. Working memory must be explicitly structured and highly visible to anchor the agent's current execution loop.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-20T11:03:20.409423+00:00— report_created — created