← sources

Agents always forget!

source: reddit (reddit:ClaudeCode)published 2026-11-11 22:05ingested 2026-11-12 02:42

retained text

    Hey all,
    
    
    I'm a solo founder, not a programmer, and I run my whole project through Claude Code. Usually 4-10 sessions at once, each with a role: one coordinator, one that merges, infra, a few workers. 1M context, Opus mostly.
    
    
    The thing that's killing me: the sessions forget. Not code, but what I told them. Tasks, decisions, "by the way do this too" stuff. It gets much worse when I load one session with lots of different small tasks, and every time a session gets full and I move to a fresh one.
    
    
    We actually measured it. On a normal evening, about 2.8 things get dropped per session switch. In one evening I found 38 dropped items. Every single one of them existed only as free text in the chat. Zero of the dropped ones were rows in our task queue. So the pattern is clear, the fix isn't.
    
    
    What we've tried so far:
    
    
    - A notes file per session that it must update every step (open items, what it's waiting on)
    - Switching to a new session at ~300k tokens with a written handoff brief, no /compact (because everyone says compact loses things)
    - A task queue CLI as the "single source of truth", every promise is supposed to become a row
    - A pile of hooks: can't end a turn with "waiting on X" unless X has a queue id, reminder after 15 turns without writing notes, PreCompact hook that dumps state, etc. Fun fact: 4 out of 4 new hooks passed their own tests and were wrong when run against real session data
    - Compact instructions in CLAUDE.md
    - Tested /compact directly: 10 explicit facts survived 8 compactions, 10/10. But in real work it still lost things. Tonight two real sessions got compacted: after re-reading their notes file nothing was missing, but from "memory" alone about 1 in 9 items was gone
    
    
    Right now I have the agents replaying 100 of my real messages from one day through sandboxed test sessions, one with auto-compact, one with handoff, one with no hooks at all, to count what each drops. Honestly I've stopped trusting the analysis, this is attempt number four.
    
    
    So, questions for anyone running Claude Code heavily:
    
    
    1. Has anyone actually solved this, or is everyone just living with it?
    2. Compact or handoff to a new session? What really loses less for you?
    3. Does "one session per task" fix it, or does the forgetting just move to whoever coordinates?
    4. What do you use as external memory that the model actually keeps up to date without being nagged?
    5. How do you check that nothing was dropped? That's the part I can't crack. I only find out days later.
    
    
    Any setup, hook, habit or tool that worked for you, I'd love to hear it. Thanks!

document fetched 2026-11-12 00:41 · canonical https://reddit.com/r/ClaudeCode/comments/1xtv55v/agents_always_forget