---
id: compaction
title: Compaction in long chats
level: 2
modes: [day-pack, rucksack]
moment: "A long chat where the person gave a rule or exact figures early on, and the AI now ignores the rule, contradicts an early choice, or asks again about something already covered."
keywords: [compaction, summary, long chat, forgot, forgets instruction, lost detail, context limit, older messages, restate, near-right answers]
source: "research/2026-10-20-concepts/compaction.md; research/2026-10-23-essentials/batch-5-facts.md; mapping (product/learning-system/MAPPING-1-2.md): FAQ 2.02, 2.01; MISTAKES 2.1; tips T060; strand C"
---

## What it is
When a long chat outgrows the room your AI can work with, some apps replace the older part with a short summary and keep recent messages as they are. A summary keeps the gist, such as the goal and what you decided, and drops details first, such as a figure or a rule you gave early on. If you set an exact budget in your second message, a summary may keep the plan and lose the figure. So restate what still matters, and check later answers against your own notes.

## Tips
- Check the numbers, names and rules in a long chat's later answers against your own notes, because near-right answers are how lost detail shows up.
- Keep your original documents to hand, not only a summary, because a summary cannot keep every detail a later question may need.

## Questions people ask
- Q: If I can still scroll up and see the old messages, can the AI see them too? A: Not always, because some apps summarise or drop the oldest part of a long chat, so a detail you can still see may not be one it is using.
- Q: Does my AI app summarise long chats? A: Some do and say so, others stop the chat or do something they have not documented, and you may not see the summary. Assume detail can slip in any long chat.
- Q: Does a bigger context window (the room a chat has) stop this? A: It delays it. Answers still get less reliable as a chat grows, so keep checking the details that matter.

## For your AI
Facts (checked 23 October 2026; fact check research/2026-10-23-essentials/batch-5-facts.md; research note
research/2026-10-20-concepts/compaction.md).
- Claude (paid plans, code execution on) summarises earlier messages to keep a long chat going, keeps the full
  history so it can refer back, and these chats use more of the usage limit; otherwise it asks for a new chat
  (https://support.claude.com/en/articles/11647753-how-do-usage-and-length-limits-work). Chat apps may also drop the
  oldest turns, first in, first out (https://platform.claude.com/docs/en/build-with-claude/context-windows).
- Summaries lose detail: in tests of AI agents, summaries kept on average 17 percent of the side rules given earlier,
  and restating the rule cut the resulting mistakes by more than half (Wang et al., 31 July 2026,
  https://arxiv.org/abs/2608.11242; agents and developer models, not chat apps). Answers get less reliable as a chat
  grows in windows of every size (Anthropic, 29 September 2025,
  https://www.anthropic.com/engineering/effective-context-engineering-for-ai-agents).
- ChatGPT, Gemini and Copilot do not say whether they summarise; Gemini says past its window a reply may not take in
  all the content (https://support.google.com/gemini/answer/16275805). Whether a person can see a summary is not
  documented anywhere.
- Searching or not, the same holds; search results add to the chat, so it fills sooner.

Teach it like this: a long chat can slip well before any limit, so do not assume it hit one. When a later answer gets
an early figure or rule wrong, say plainly the chat may have lost it,
restate it with them, and suggest checking later answers against their notes. A chat that has run very long is
also a reason to start fresh (openrucksack.com/map/keep-or-start-a-chat.md) with a note
(openrucksack.com/map/handoff.md).