Anthropic patches 'forgetfulness' bug that erased Claude’s internal thought traces
- Anthropic said on April 23 that three changes degraded Claude Code, including a March 26 bug that made resumed sessions keep deleting older thinking. - The bug was meant to clear stale thinking once after an hour idle, but instead fired every turn until Anthropic fixed it April 10. - It matters because coding agents can look weaker from scaffolding bugs alone, even when the underlying model and API still work.
Claude Code didn’t suddenly get dumber. That’s the core of this story. What broke was the software around the model — the harness that manages long coding sessions, cached context, and how much reasoning the assistant does by default. Anthropic said on April 23 that three separate changes degraded Claude Code, Claude Agent SDK, and Claude Cowork over several weeks, and one of them made Claude seem weirdly forgetful after idle sessions. (anthropic.com) ### What actually went wrong? The headline bug was a session-management mistake. On March 26, Anthropic shipped a change meant to clear older “thinking” from sessions that had been idle for more than an hour, basically to make resumed sessions faster. But the bug meant that cleanup didn’t happen once. It kept happening on every turn for the rest of the s(anthropic.com)look repetitive, forgetful, and unable to build on its own prior work. Anthropic says it fixed that on April 10. (anthropic.com) ### What does “older thinking” mean here? This is not the model erasing its weights or forgetting training. It’s much narrower than that. Claude Code keeps session state — including prior reasoning and working context — so it can continue a long task across many turns. If the system keeps trimming that state incorrectly, the model still runs, but it star(anthropic.com)sult is not total failure. It’s worse — a kind of plausible incompetence. (anthropic.com) ### Why did users feel it as “quality got worse”? Because from the outside, infrastructure bugs and model regressions can look identical. If Claude forgets what file it was editing, repeats steps, or loses the thread in a debugging session, most users will call that a weaker model. Anthropic says the API and inference layer were never affected, which is (anthropic.com)t wrapper was. But if you use Claude Code as a coding partner, that distinction barely matters in the moment. Broken behavior is broken behavior. (anthropic.com) ### Was this only one bug? No — and that’s why the whole episode looked messy. Anthropic says there were three separate issues. First, on March 4 it lowered Claude Code’s default reasoning effort from high to medium to reduce long pauses and frozen-looking UI. Then came the March 26 thinking-history bug. Then on April 16 it added a prompt instruction to (anthropic.com)at hurt coding quality and got reverted on April 20. All three were resolved by April 20 in version 2.1.116. (anthropic.com) ### Why was this hard to diagnose? Because each change hit a different slice of traffic on a different schedule. Anthropic says that made the aggregate effect look like broad, inconsistent degradation rather than one clean failure mode. That matches how users described it — some saw shallower reasoning, some saw forgetfulness, some saw terse answers that(anthropic.com)gger theories about silent model downgrades or overloaded servers. Anthropic explicitly said it does not intentionally reduce model quality due to demand or time of day. (anthropic.com) ### Why does this matter beyond Claude? Because coding agents are not just models. They’re models plus memory handling, tool loops, prompts, permissions, caches, and UI decisions. Anthropic itself has written that agent performance can vary a lot based on the surrounding scaffolding, even with the same underlying model. So this incident is a reminder tha(anthropic.com) model itself. (anthropic.com) ### Did Anthropic do anything for users? Yes. Anthropic said on April 23 that it was resetting usage limits for all subscribers after resolving the issues. That’s a practical acknowledgment that people burned time and quota on degraded behavior, not just a vague apology blog post. (anthropic.com 1) (anthropic.com 2)y complex product does. It’s that a bug in session-state handling could make a top-tier coding model look flaky, forgetful, and inconsistent for weeks. That’s where AI products are now — the frontier is not only raw intelligence, but whether the system can hold onto its own train of thought long enough to be useful.
