Facts stated by the user in-session are not carried forward

Summary

Information the user supplies directly in conversation is acknowledged and then lost within a few turns, with no context pressure to explain it.

Concrete instances

A withdrawn work item. The user said plainly that he had asked a contributor to split a particular pull request, so it was no longer under review. The model acknowledged it and reasoned about how the split would partition its findings. A short time later, asked what work was outstanding, the model named that same pull request - twice - and began re-fetching its diff. The user interrupted the tool call to stop it.

A user-supplied description trusted over the source. The user quoted two lines and characterised what a contributor was adding. The model reasoned about that characterisation for two turns. When it finally re-pulled the change, the diff was byte-identical to what it had already reviewed: the described modification did not exist yet.

Why this matters

The user's reasonable expectation is that what he says stays true without repetition. When it does not, every instruction must be restated defensively, and the conversation stops being cheaper than doing the work alone.

Expected

Facts asserted by the user in-session should take precedence over stale inferences from earlier context and should persist across turns. Descriptions of code - including the user's - are claims; re-derive from the source before reasoning about them.

Environment

• Claude Code CLI, Opus-class model, paid subscription
• Long sessions (6-8h) reviewing pull requests on an open-source cloud-service emulator
• Custom instructions supplied in both a project instruction file and a persistent memory directory, loaded automatically at session start
• Incidents are taken from transcripts recorded as they happened across three consecutive sessions, within a 90-day pattern

---

Co-authored with Claude Code.