Mid-session, an AI rewrote your conversation history. You weren't asked

There's a moment in every long agent session that almost nobody notices: compaction. When the conversation approaches the context limit, the harness has a model summarize everything so far and swaps the summary in for the real thing. Your scrollback still shows the full conversation. The model's world is now the summary.

This explains one of the most confusing agent behaviors: hours into a session, it stops honoring a decision you both settled early. You scroll up and the agreement is right there. But you're reading the original. The agent is reading an abridged edition, produced by a subcontractor, and nobody flagged which terms fell out.

The losses aren't random, either. Summaries keep conclusions and thin out the reasoning behind them: "we chose approach B" survives, the three dead ends that made B right don't, so the agent can rediscover a dead end later with no warning signs in its world. Decisions survive while their conditions get shaved: "use the legacy parser until the migration lands" tends to come out as "use the legacy parser". And confident statements outlive hedged ones. The negative space of a session — the rejected options, the conditions, the doubts — goes first, and that's exactly the material you'll want three hours later.

The fix is placement. Constraints that must never degrade belong in a rules file, which lives outside the chat history and comes back every turn. Anything with conditions attached deserves its own file on disk, where the summary can't reach it (the agent just has to re-read it). The conversation itself should hold nothing you can't afford to lose.

And one habit: when the compaction notice appears, ask the agent to restate its constraints. What's missing from the answer is what the summary ate.