Amp July 2, 2026 Changelog

Amp is an AI coding assistant, and this update is all about how it reads its own past conversations. This matters to anyone who works on long-term projects, because it means you can now ask Amp to look back at older, massive chat threads and actually get accurate answers instead of confused summaries.

What's new

• Smarter thread reading: Instead of trying to read a giant conversation all at once, Amp now uses a dedicated "subagent" (a smaller AI helper) to search through long threads, read the specific messages, and answer your questions.
• Checks its own history: The AI can now look back at the current thread you're working in. If you need to know a decision from three weeks ago, it will actually go look for it instead of just guessing based on a summary.

Improvements

• Better accuracy over speed: Amp used to stop searching as soon as it found the first answer, which was often an old, failed attempt. It's now programmed to keep searching to see if newer messages changed or reverted that first attempt.
• Verifies actions: Amp now understands that just because it *tried* to do something in the past doesn't mean it actually worked. It checks to see if an action actually succeeded before believing it happened.
• Handles "compaction" better: When conversations get too long, Amp compresses them to save space (called compaction). It now uses these summaries for general context, but will look at the original messages when you need exact code or details.

Fixes

• Fixed the "middle of the conversation" bug: Before, if a thread was really long, Amp would only pay attention to how it started and how it ended, completely ignoring everything in the middle. The new update fixes this so the whole conversation is searched.
• Fixed crashes on huge threads: Previously, if a thread got too long (like millions of words), it would crash the system that tried to read it. The new searching method prevents this.

Breaking changes

No breaking changes in this release.