Your agent isn't hallucinating, it's working from stale notes
retained text
One comment from these threads stuck with me: the hard problem isn't remembering, it's updating what the agent thinks it already knows. That's the failure mode once you're running agents overnight. Memory is the easy half. A notes file or a CLAUDE md mostly fixes that. Staying current is the half that still hurts. The notes pile up, the agent treats them as ground truth, then redoes work you already finished because a line in the status file said it still needed doing. What I've been trying: keep the status note short and dated, have each session write what it changed and what it did not get to so the next one reads a handoff instead of a knowledge base, and treat every note as a claim until you check the file or the git log. Prune on purpose. Most of that is discipline, and discipline is the first thing to go past three sessions. I've been building a Kanban app so the handoff lives on the task instead of a growing notes file. It doesn't fix everything, especially how much context you still load in. For people running this daily: do you prune your agent notes, or let the file grow and hope the agent re-checks the code before trusting it?
document fetched 2026-11-08 00:33 · canonical https://reddit.com/r/ClaudeCode/comments/1xpsqlu/your_agent_isnt_hallucinating_its_working_from