Every time you open Claude today, it starts from scratch. It has no memory of your job, your projects, your preferences, or anything you told it last week. That is not an accident; it is a design choice Anthropic made in the name of user control and privacy. But that choice has a real cost, and Anthropic is now moving to change it.
The incoming feature, called Memory Files, would split what Claude knows about you into separate documents organized by topic, project, or context. Today, Claude compresses whatever it learns into one running summary. That summary gets unwieldy fast and has hard limits on how much it can hold. The file-based approach lets Claude pull in only the notes relevant to the current conversation, which keeps things clean and means the memory can scale without becoming a bottleneck.
What makes this worth paying attention to is the user control piece. You will be able to open, read, edit, and delete these files whenever you want. That is a different design philosophy from ChatGPT, which runs two memory layers simultaneously: a profile of saved facts you can see and inspect, and a second layer on paid plans that references patterns from your full conversation history, sitting outside that visible list. The result with ChatGPT is convenience alongside some opacity. Anthropic is going the other way, making every stored record visible and editable by design.
Tied to Memory Files is a background process called Dreams. It runs on a schedule when you are not actively using Claude, scanning accumulated memory files to merge duplicates, remove outdated notes, and surface patterns Claude missed during live conversations. Anthropic compares this to the way the brain consolidates memories during sleep, with original files left untouched while a cleaned-up version is prepared. Dreams was announced at Anthropic's Code with Claude developer conference and is currently in beta for developers, scoped to specific model versions. It has not yet reached the general consumer product.
The broader context involves an internal Anthropic project called Conway, which surfaced when Anthropic accidentally published around 512,000 lines of Claude Code source to a public npm package on March 31, 2026. Anthropic confirmed the incident was a release packaging error caused by human error, not a security breach. Developers who examined the leaked code found references to Conway: an always-on agent environment built to run in the background, respond to external triggers like an incoming message, and execute tasks without waiting for you to open a chat window. Anthropic has not confirmed Conway publicly or commented on what the code revealed.
What the Conway discovery does clarify is why persistent, organized memory matters so much right now. An agent that runs in the background and acts on your behalf is essentially useless without a reliable record of who you are, what you are working on, and what your preferences are. Memory Files looks like the foundation that kind of product needs.
There is also a gap worth noting for anyone thinking about team use. Memory in Claude today is personal only: each user maintains their own separate record, with no mechanism to share a common memory layer across a group. That matters for organizations that want an AI assistant with institutional knowledge, not just individual notes. No shared memory feature has been announced, and no timeline exists for one.
No firm launch date has been announced for Memory Files in the consumer product. The direction is clear; the pace is not.