Anthropic Plans Claude Memory Overhaul with File-Based 'Memory Files' System
Summary
- • Anthropic preparing 'Memory Files' to replace Claude's single-note memory model
- • File-based system organizes user context across topic- and project-specific documents
- • Async 'Dreams' feature consolidates memory between sessions, inspired by REM sleep
- • Overhaul likely foundational to upcoming Claude Conway agent; no release date set
Details
Memory Files replaces single-summary model with multi-document architecture
The current 'classic' system compresses everything learned about a user into one rolling note. Memory Files splits this across multiple discrete documents per topic, project, or context — allowing Claude to load only relevant files into its context window rather than ingesting an entire summary that may contain irrelevant or outdated information.
'Dreams' performs async memory consolidation between sessions without touching originals
Dreams is a scheduled background process that merges duplicate entries, replaces stale values, resolves contradictions, and surfaces patterns missed during live sessions. It produces a new reorganized memory store rather than modifying the original in place — Anthropic compares the process to REM sleep memory consolidation in humans.
Dual-mode design lets users choose between classic and Memory Files architectures
The phased transition approach reduces friction for existing users. Users will be able to browse and edit their Memory Files directly, aligning with Anthropic's stated emphasis on user transparency and control over retained data.
Dreams currently in limited beta scoped to Opus 4.7 and Sonnet 4.6 on developer platform
The Dreams feature is not yet in the consumer Claude product — it remains in restricted beta on Anthropic's developer/managed agents platform. No general availability date has been announced for either Memory Files or Dreams.
Memory Files mirrors filesystem-style memory already used by OpenClaw and Hermes agents
Third-party agentic systems built on Claude — specifically OpenClaw and Hermes — already use filesystem-style memory to overcome the limits of rolling summaries. Memory Files brings this pattern natively to Claude, and may be an evolution of a previously discovered internal feature called 'Knowledge Bases.'
Update likely foundational for forthcoming Claude Conway agent
Conway is described as arriving soon, and the structured, scalable memory architecture would logically underpin an agent requiring richer, longer-horizon user context — suggesting Memory Files is part of a broader platform architecture shift, not just a standalone feature.
Product Launch = new capability announcement, New Tech = novel technical mechanism, Strategy = design or deployment decision, Tech Info = current scope and limitations, Context = background or prior art, Industry Update = broader platform implications
What This Means
Persistent, structured memory has become a key competitive battleground in the consumer AI assistant market, and Anthropic's Memory Files push Claude toward parity with rivals that have already deployed long-horizon memory systems. The Dreams consolidation layer is a technically interesting approach to a real problem — memory stores that accumulate noise, drift, or contradiction over time — and the sleep-cycle analogy suggests Anthropic is thinking about memory as an ongoing maintenance problem, not just a storage problem. If the user-editable, file-based approach lands well, it could set a new standard for how AI assistants handle personal context: transparent, auditable, and modular rather than opaque and monolithic.