Developer Murooka’s tests provide a concrete figure: after three consecutive compactions in Claude Code, two real development conversations lost 33% and 57% of user instructions, respectively. Asking the summary to preserve historical messages verbatim reduced the loss rate to zero, but extremely long sessions still produced omissions and truncations.
This issue matters because conversation compaction in Claude Code—the process of replacing a long conversation with a summary to conserve computing resources—is not an occasional glitch. It may become a systemic risk for AI coding tools that run continuously for several hours. Prompts can mitigate the immediate problem, but they cannot replace disciplined rule management.
What this is
Claude Code automatically compacts a session when it becomes too long, and users can also trigger it manually with /compact. After compaction, earlier conversations are replaced by a summary. Earlier constraints such as “use Japanese,” “fix the bug first,” or “do not change the existing design” may disappear permanently if the summary merely paraphrases them.
Murooka ran three rounds of testing on two real conversations involving VS Code extension development. The first compaction lost none of the 18 sampled instructions, but by the third consecutive compaction, the loss rates had risen to 33% and 57%. At the same conversation length, a single compaction lost only two of 54 instructions, indicating that the problem primarily comes from “summarizing a summary,” not from insufficient message length.
He added a prompt after /compact requiring user messages from the previous summary to be copied verbatim into the new summary. Across three consecutive test rounds, this reduced the instruction loss rate to zero.
Industry view
We believe this exposes a contradiction that AI coding tools have yet to fully resolve: long-running tasks require continuous context compaction, yet compaction is itself the main source of information loss. The longer a tool can operate without interruption, the more summary layers it accumulates—and the more likely errors are to compound.
This method should not be treated as a universal fix. Testing showed that once a session accumulated hundreds of messages and tens of thousands of characters, asking the model to copy the original text still resulted in omissions and truncations. Large volumes of verbatim text also crowded out other material from the summary. The test did not verify whether Claude actually followed the instructions retained after compaction.
A more reliable approach is to put long-term rules in CLAUDE.md, the project-level rules file read by Claude Code, rather than relying on chat history. Claude Code’s official documentation also warns that detailed early instructions may be lost during compaction and recommends placing long-term rules in this file. The risk is that users may not recognize which requirements constitute persistent constraints; the tool also does not automatically audit whether an organization’s rules are complete.
Impact on regular people
For enterprise IT: When deploying tools such as Claude Code, encode critical business constraints, coding standards, and acceptance criteria in project rules or tests rather than relying solely on prompts. Teams should also establish checkpoints for long-running tasks to reduce accountability gaps caused by repeated compaction.
For individual professionals: When working with AI on tasks that run for several hours, do not assume that “it remembers everything.” Record important decisions in documents, task lists, or code, and periodically verify that critical constraints are still being enforced.
For the consumer market: As AI assistants and coding agents move from short question-and-answer sessions toward extended execution, the reliability of summaries will directly affect renewals and user trust. If vendors showcase task-completion speed without explaining how context is preserved, they may widen the gap between user expectations and the actual experience.