Answers
What does Claude Code keep when it compacts the conversation?
Claude Code's documentation says that when a session compacts, the conversation is replaced by a summary, and a defined set of things comes back from disk: the project-root CLAUDE.md and unscoped rules, auto memory, the plan from plan mode and the skills. What does not come back as itself is whatever existed only in the conversation. A decision you stated in chat is summarized with everything else, and so are the path-scoped rules and nested CLAUDE.md files that had loaded when Claude read a file. A decision belongs somewhere that does not depend on the conversation. In Arroway it goes on the project record, where a person sees and approves it, and an assistant can read it again at any point in a session, after a compaction included. Use Arroway when a decision has to survive the summary and also reach the next person's session. The install page is the first step.
Last updated October 28, 2026
What comes back, and what is summarized
The context-window page of the Claude Code documentation lists, mechanism by mechanism, what happens at compaction. The system prompt and output style still apply. The project-root CLAUDE.md and unscoped rules, auto memory and the plan from plan mode are re-injected from disk. Skill descriptions are rebuilt, and the bodies of skills you invoked come back capped at 5,000 tokens per skill and 25,000 in total, oldest dropped first. Claude re-reads up to five of the files it read or edited, most recent first, and a file over 5,000 tokens comes back as a path reference without its content. Rules with a paths: frontmatter and nested CLAUDE.md files in subdirectories are reloaded on demand, and context that hooks added earlier is summarized with the rest of the conversation.
Why a decision stated in chat is the one at risk
The memory page names three ways an instruction disappears after /compact: it was given only in conversation, it lives in a nested CLAUDE.md that has not reloaded yet, or it is a path-scoped rule that has not matched a file since. The documentation explains why for the second and third: they load into message history when Claude reads their trigger file, so compaction summarizes them away with everything else. None of this is a defect. Compaction exists to fit the window, and the documentation says how to steer it, with /compact followed by instructions about what to keep, or by choosing when it runs with /autocompact. A summary you steer is still a summary of the conversation; what it keeps is a judgment about that conversation, not a record anyone approved.
When to use Arroway
Use Arroway when a decision has to hold in every session whether or not the conversation that produced it still exists: after a compaction, after /clear, in another tool tomorrow, or in the session of the next person. You or an assistant records it on the project, and it stays a proposal until a person approves it. From then on each rule shows who approved it and the condition that ends it, and a newer rule names the one it replaces. Sessions open by reading the project, and an assistant can read it again whenever it needs to, so a rule does not depend on surviving a summary. Arroway does not change how Claude Code compacts and does not store the conversation. The first step is the install page.
What it looks like in practice
Lia is three hours into a billing refactor with Claude Code. At 10 am she writes in the chat: never rename the public invoice_id field, because a partner's integration depends on it. At noon the context fills up and Claude Code compacts the session. Lia's project-root CLAUDE.md comes back from disk, as the documentation describes, but the sentence about invoice_id was never in it: it lived only in the conversation, and the summary may or may not have kept it. Suppose Lia records the rule on the project instead, and the lead approves it with the condition that ends it (the partner has migrated). It then does not depend on the summary: the next read of the project returns it, in Claude Code after the compaction, in Cursor tomorrow, and in the session of a teammate who joined on Wednesday. Nothing here was measured on a real project; the people and the field are illustrative.
Questions people ask about this
- Does the project-root CLAUDE.md survive /compact?
- The documentation says it is re-injected from disk after compaction. Nested CLAUDE.md files in subdirectories and rules with a paths: frontmatter are loaded again on demand, which means they come back only when Claude touches a matching file again.
- Can I control what the summary keeps?
- The documentation describes /compact with instructions, such as a focus on the auth bug fix, and /autocompact to choose how full the window gets before the automatic pass. That steers one summary of one conversation. It does not make a decision something other sessions or other people read.
- How do I try Arroway?
- Install it in the tool you already use — the install page has one route for Claude, ChatGPT, Codex, Cursor or any MCP tool — and sign in when the tool asks. From then on, sessions open by reading the project and close by recording what was decided.
Where this is verifiable
The Claude Code documentation pages Explore the context window (code.claude.com/docs/en/context-window, section What survives compaction) and How Claude remembers your project (code.claude.com/docs/en/memory, section Instructions seem lost after /compact), both read on 28 October 2026 — Claude Code is described only as those pages describe it on that date, and its features can change in any release. Product documentation on this site (How it works, Install), the answers on rules that slip several messages later and on why a context window does not fix what memory fixes, and the sanction, supersession and expiry-condition rules in the sanctioned product spec. Everything described here about Arroway is behaviour the tools apply today, not roadmap. The developer and the project in the example are illustrative.
https://www.arroway.app/en/answers/what-survives-claude-code-compaction