Claude is about to get a real memory — and this is bigger than it sounds

🍀

May 25 • AI news

Quick gut check: how many times have you re-explained the same context to Claude? Your business, your ICP, your stack, your preferences… over and over, every new chat.

That's about to change — and the way they're doing it is the part worth paying attention to 👇

Anthropic is testing a dual-mode memory system. You'll get to pick between two setups:

Classic — what we have now. Everything Claude learns about you gets compressed into one summarized note.

Memory Files — the new one. Your context gets split across multiple structured documents, organized by topic, project, or context. Notes Claude writes as you chat and reads back only when they're relevant — and you can browse and edit them anytime.

Basically a built-in personal wiki Claude pulls from selectively, depending on what you're working on.

Tied to this is a feature called Dreams, which Anthropic has already shipped (in preview) on its Managed Agents platform. It's a scheduled, background pass over your memory — merging duplicates, replacing stale info, resolving contradictions, and surfacing patterns it missed live. They literally model it on REM sleep: your brain replaying the day and deciding what to keep. The original store stays untouched while a cleaned-up version gets produced for review.

Why this matters if you're building with AI:

If you've touched always-on agents, you already know this pattern — file-based memory is what lets agents scale past the limits of a single context window. Anthropic is now bringing that same architecture toward the consumer product. That's the real signal here. They're not just improving chat memory — they're laying groundwork for persistent, always-on agents.

For builders: this is the unlock for agents that actually remember the job. No more cramming everything into a system prompt and praying. You'll architect memory the way you'd architect a knowledge base — by project, by client, by context. (Side note: Harvey, one of Anthropic's enterprise customers, reportedly saw task completion rates jump ~6x once Dreams-style consolidation was turned on. That's not a rounding error.)

For operators: think about what this does to a Claude running your business workflows. It stops being a tool you re-brief every morning and becomes something that genuinely knows your context — your CRM quirks, your team, your SOPs — and pulls the right piece at the right time.

The detail I love most: user control stays front and center. You can browse and edit the files. That directly answers the #1 trust objection I hear — "what is this thing actually storing about me and my business?" You get to see it. You get to edit it. That's the difference between a memory feature people fear and one they'll actually switch on.

No firm consumer timeline yet, and the deeper agent tie-ins are still reading-the-tea-leaves territory. But this is the most consequential thing in Anthropic's pipeline right now — and it puts Claude on real competitive footing with the persistent-memory plays everyone else is chasing.

When it drops, I'll show you exactly how to structure your memory files for an AI business workflow. Going to be a game-changer for anyone running agents.

Team Classic or team Memory Files? I already know where I'm landing 👇

8 comments

Claude is about to get a real memory — and this is bigger than it sounds