Claude’s New Dual‑Memory System: Is a ‘Permanent Brain’ Finally Here?

Dual‑Memory Architecture

Anthropic is testing a new "dual‑memory system" for Claude. One side retains the existing "classic memory," a single rolling summary that compresses all user information into one abstract. The other side introduces "Memory Files," a persistent, file‑based storage that behaves like a personal Wiki.

When activated, users can switch freely between the two modes, marking the most extensive overhaul of Claude’s memory since its launch.

Memory Files – Unlimited, Structured Storage

Classic memory overflows when the information volume grows, causing newer details (e.g., today’s coffee preference) to overwrite older, more important context (e.g., a month‑long product‑architecture discussion). Memory Files replace this by storing each topic, project, or context as a separate, structured document that can be retrieved on demand.

Benefits highlighted by the article:

Capacity ceiling broken – file‑system‑based management can theoretically expand without the length limits of a single summary.

Precision improves exponentially – during code discussions Claude pulls only technical docs; during travel talks it pulls only travel preferences, avoiding the brute‑force injection of all memories.

User regains control – users can browse, edit, or delete any memory file like a Wiki entry; unwanted memories are removed simply by deleting the corresponding file.

Dreams – The REM‑Style Background Consolidation

The "Dreams" preview, inspired by human REM sleep, runs asynchronously after a period of inactivity. It replays daytime experiences, strengthens important connections, discards noise, and merges short‑term traces into long‑term memory.

Trigger conditions: at least five dialogue turns, or more than 24 hours since the last consolidation, or manual activation via the /dream command.

During a Dream cycle the system performs four operations:

Merge duplicate entries across files.

Replace outdated items (e.g., "we decided on Redis" becomes "2026‑05‑15 we decided on Redis").

Resolve logical contradictions by keeping the newer, more reliable entry.

Surface hidden patterns that neither human nor AI noticed during the conversation.

Conway – A 24/7 Autonomous Agent

Alongside Memory Files and Dreams, Anthropic introduced "Claude Conway," a continuously running AI agent that never goes offline. Conway provides three core modules: Search, Chat, and System, and can listen to external events, trigger tasks, receive webhook signals, control browsers, and execute Claude Code.

Conway runs on Anthropic’s hosted cloud infrastructure, offering tighter security than open‑source agents like OpenClaw, which suffered multiple CVE disclosures within its first two months.

Industry Context and Competitive Landscape

The article cites TestingCatalog as the source of the leak and notes early adopters such as Netflix, Rakuten, and Wisedocs, which reported a 97 % drop in error rate and a 30 % speed‑up in document verification after using the new memory system.

It contrasts Claude’s approach with competing strategies: OpenAI’s upcoming "Memory Sources" (splitting memory into saved memories, chat history, custom instructions, file libraries, and Gmail) and Google Gemini’s deep integration with personal data (Gmail, Drive, Calendar). Claude’s path combines structured storage, autonomous consolidation, and an always‑on agent, forming a closed loop from memory to reflection to action.

Finally, the article argues that persistent, scalable memory is likely a necessary component for future artificial general intelligence (ASI), echoing Dario Amodei’s view that ASI will emerge from the gradual assembly of modular capabilities.