Claude Code is quietly wiping out older chat transcripts, and the default timeline is blunt: 30 days. Users say the coding conversations they treated as reusable working knowledge can vanish the next time they open the app, and the cleanup seems to happen without any meaningful first-run warning.

The mechanism at the center of the complaints is an option in Claude Code called cleanupPeriodDays. It defaults to 30 days and runs every time Claude Code starts up, deleting any.jsonl file that is not fresh enough. In other words, the product is not just “forgetting” in the abstract. It is running a file cleanup job on disk, and users only discover it after they notice missing transcripts.

This might be more tolerable if Claude Code told users what it was doing. But based on the GitHub issue threads referenced in The Register’s report, multiple users say there was no install-time disclosure and no first-run dialog explaining that their chat transcripts would be erased on a schedule. One GitHub user, FTSBrand, wrote that cleanup runs out of the box with no install-time disclosure or first-run dialog, and that users who treat conversation history as durable working knowledge are silently mistaken about the persistence model. Another user reports that some project artifacts can remain after cleanup, but the “reasoning trail” and design, debugging, and analysis context are gone.

Anthropic, the company behind Claude Code, pushed back on blame for misunderstanding. The Register reports that Anthropic suggested the issue is rooted in users not checking documentation, and it said the 30-day erasure policy has been there since Claude’s launch as a security measure and is documented. In a statement quoted by The Register, Anthropic said: “Keeping plain text transcripts of coding sessions on disk indefinitely creates real security and privacy risks, since they can contain source code, credentials, and other sensitive material.” The company added: “The 30-day default balances the ability to resume recent work against not holding that data on disk longer than needed. This has been part of Claude Code's design since launch as a security measure.”

From a board and compliance lens, the statement is easy to understand, but it lands with a different kind of risk when the product behavior is not clearly surfaced at the moment it matters. If a tool can delete potentially sensitive local artifacts on a predictable schedule, then internal stakeholders need to know two things immediately: first, when deletion occurs, and second, whether the retention setting behaves reliably. When those details are not apparent to users, the “security measure” can turn into an operational incident. You do not only need to protect data from unauthorized access. You also need to ensure teams can prove what happened, when it happened, and whether a retention change actually changes the outcome.

That “prove it” requirement shows up in the complaints about recovery and auditability. The Register reports that users say cleanup appears to bypass recovery, with no soft-deletion option, grace period, or option to restore. Users also suggest there is no log of what gets deleted, leaving them unable to confirm what was wiped after it happens. And the story gets more complicated when users try to adjust retention. Some people set cleanupPeriodDays to a large value, expecting it to prevent deletion, but report that it is not working properly.

One GitHub user, ojura, provided root-cause analysis in the thread, according to The Register. The claim is that deletion is keyed to a transcript’s mtime, meaning modification time, rather than its actual last activity timestamp. Because mtime is externally mutable, ojura argued, anything that touches it can flip the outcome: a restore, a sync client, or a script that sets mtimes to a session’s true (old) last-activity date makes a present session look old, and it is silently deleted on the next sweep. That matters because it means “I changed the setting” is not the same as “I changed what the system will do.” It also means that backup and sync tooling, which organizations commonly run for developer workflows, could accidentally interact with the cleanup logic.

The most consistent solution described in the thread is to back up Claude Code transcripts. The Register notes multiple workaround iterations, including ensuring transcripts are backed up. But backups, as one GitHub user, caioribeiroclw-pixel, wrote in a message quoted by The Register, are “good hygiene” and do not replace product-level disclosure and provenance for a destructive retention sweep. In other words, even if the retention behavior is eventually documented, the operational reality is that developers need clarity and confidence at the time they click “use.” For executives, the second-order implication is that AI tooling is no longer just a performance layer. It is joining the stack of systems that handle sensitive intellectual property and credentials, and it behaves like a data lifecycle system whether or not it looks like one.