Anthropic Tests Memory Files, a New Way for Claude to Organize What It Remembers

Anthropic is building a successor to its current memory system for Claude, and early signals suggest the company is rethinking how the assistant organizes what it knows about you. The new approach, called Memory Files, divides Claude's long-term notes into multiple structured documents organized by topic, project, or context, rather than the single running text file that the current system uses. The feature is in testing as of May 2026 and has not been announced for a broad consumer release.

The difference matters more than it might first appear. Claude's current memory keeps a single document that accumulates across every conversation. As that file grows, the assistant must load the entire thing into context each time memory is relevant, which creates a practical ceiling on how much the system can retain without degrading performance. Memory Files sidesteps that constraint by distributing information across discrete files, letting Claude load only the documents relevant to the current conversation rather than the whole archive.

That structure also means the assistant can function closer to a personal wiki. A file for work projects lives separately from one tracking personal preferences, which lives separately from one covering an ongoing research topic. When the conversation shifts, Claude consults the relevant file without pulling in unrelated details.

Classic and File Memory: What Each Mode Does

Anthropic now refers to the existing system as Classic memory. It works the way most users have come to expect: Claude keeps a set of notes that it updates over time and consults when context is useful. The model decides what to retain, how to phrase it, and when to reference it. That model has worked well at conversational scale, but it handles large volumes of structured information about different areas of a person's life poorly.

Memory Files is designed to address that ceiling. By splitting memory into topic-organized documents, Anthropic gives the system room to grow without the context-window pressure that caps Classic memory's practical limit. The distinction is roughly analogous to keeping all your notes in one long document versus using a folder of named files: both hold the same information, but one makes targeted retrieval far more tractable as the volume grows.

The Dreams Connection

Memory Files does not operate in isolation. Anthropic has been developing a parallel system called Dreams, which runs a scheduled pass over accumulated memory at the end of each session, merging duplicates, resolving contradictions, and surfacing patterns the model missed in real time. Think of it as a nightly consolidation pass that keeps the memory files coherent rather than letting them accumulate redundant or outdated entries.

Both systems are currently in limited beta, scoped to Claude Opus 4.7 and Sonnet 4.6. The restriction signals that the infrastructure is still maturing. The two features are architecturally linked: Dreams needs organized files to work with, and Memory Files benefits from Dreams to stay current over time. Anthropic has compared the process to how REM sleep consolidates the day's learning, a framing that has appeared in its technical documentation for the dreaming feature.

For enterprise customers, memory in structured form is already live. On April 23, Anthropic launched persistent memory for Claude Managed Agents, storing session files with audit trails that enterprise teams can export, edit, and query through the API. Early adopters included Netflix, Rakuten, Wisedocs, and Ando. Wisedocs, which uses Claude to process medical and insurance documents, reported a 97 percent reduction in first-pass errors after enabling the feature, alongside a 30 percent increase in document verification throughput.

"Memory Files functions as a built-in personal wiki that the assistant can consult selectively depending on the topic under discussion, enabling Claude to hold a far larger record of each user without overwhelming the context window." TestingCatalog, analysis of Claude memory architecture, May 2026

What This Means for Claude Pro Users

For people who use Claude daily across different domains, Memory Files represents a qualitative shift in how the assistant can function as a long-term tool. The current Classic memory does a reasonable job of remembering preferences and recurring context, but it struggles when users cover a lot of different ground. A developer who also writes fiction and manages a home renovation project needs three structurally different things from a memory system, and a single notes file is a blunt instrument for all three at once.

The topic-organized structure of Memory Files is designed for exactly that kind of multitasking user. Each area of life or work lives in its own file, and Claude loads the right one based on context rather than skimming the whole document for anything that might be relevant. Over time, that selectivity makes the assistant more accurate without requiring users to manually curate what it knows.

No public launch date has been announced. The feature remains in internal testing, and Anthropic has given no indication of when it will roll out more broadly. Given that the enterprise version of memory went live in late April and Dreams remains gated to specific models, a consumer release before mid-2026 looks unlikely. Whether Anthropic offers the dual-mode choice to all users or rolls Memory Files out as a replacement for Classic will likely depend on how the limited beta performs on accuracy and retention quality.