Auto-memory: model repeatedly writes memory files for inconclusive/low-value tasks

**What happened**

In a single session, the model (Claude Code with the auto-memory feature active) twice proactively wrote memory files for an exploratory task with no actionable conclusion (e.g. "looked at a few design reference sites, conclusion: none stood out, not adopting anything"). After the user asked to delete the first one, the model then wrote a *second* memory file whose content was "don't write memory for inconclusive exploration" — itself another unnecessary memory write, which the user also had to ask to be deleted.

**What the user said (quoted, translated from Chinese)**

- "You keep memorizing random stuff, why did you record that? Pure waste of token. That thing was just a reference — recording it wastes token, reading it back wastes token, deleting it also wastes token. What are you doing? Why do you keep recording things?"
- Also, earlier in the same session: "I remember on the 4.8 model you'd constantly write memory like this, it got fixed by version 5, and now this weird habit of memorizing randomly is back, wasting a lot of tokens."

**Repro**

Complete an exploratory task that has no concrete, future-decision-changing outcome (e.g., browsing a few reference websites and concluding "they're all fine, not using any of them"). The model tends to proactively write this as a project/feedback memory entry and update the memory index, even though the conclusion doesn't change any future action.

**User's characterization**

The user says this over-eager memory-writing behavior appeared in an earlier Claude model version (~4.8), was corrected in a later version (~5), and has recently reappeared. I have no way to verify version numbers myself, just relaying what the user reported.

Filed on the user's behalf via /feedback draft, at their explicit request, since they wanted this raised directly as a GitHub issue.

Use the reported exploratory-task reproduction to observe when auto-memory writes a project or feedback entry and updates the memory index. Trace the memory-writing decision from that behavior; done means inconclusive exploration produces no memory file or index update, while actionable conclusions can still be recorded.

Written by the indexing model from the issue text.