Anthropic Showed How Claude Agents Remember, and Recheck, Their Work

Anthropic showed how Claude agents remember, and recheck, their work

Most AI tools start every day from zero. Anthropic wants its agents to remember. In a research post, Anthropic laid out how Claude agents keep memory: notes written from experience, saved between sessions, and loaded back in when new work starts. The notes are plain files the agent can read, correct, and build on.

The part that matters most is the recheck. Anthropic runs a background pass it calls dreaming. While the agent is idle, dreaming compares its notes against the original transcripts of the work and fixes mistakes. In one test, a note had a location wrong, and the agent kept failing the same task in a loop. After dreaming corrected the note against the transcript, the agent went straight through.

The gains showed up on a bigger test too. Anthropic gave Claude about 350,000 company files and let agents build a linked set of notes over 128 sessions. Scores on a set of test questions rose from 13% to 48% as the memory grew. Those are Anthropic's own test numbers, not an outside check. The same post carries a warning. One model saved everything, 413 files of notes, and did worse. A later model kept nine files of tested rules and went much further.

An agent that remembers your team's work changes the math. It stops relearning the same lessons, so each week of use can make the next week cheaper and better. The memory itself is a file a team can read and audit, which is rare in AI. And the recheck is the honest part. Memory that is never checked drifts, and a confident agent working from a wrong note is worse than one that starts fresh.

Apple is pushing Siri to act more like an agent

If you use an iPhone, Siri still feels like a tool for small jobs. You ask, it answers, and the task ends there. The Verge reported on Sunday that Apple wants more from it. Bloomberg's Mark Gurman says chief executive John Ternus is pressing AI chief Craig Federighi to speed up Siri's roadmap.

Two upgrades are expected in the coming months. The first gives Siri closer control of Apple's own apps, like rotating a photo in Photos or sorting tabs in Safari. That step is expected with iOS 27.3. The bigger step comes later, likely in the spring. Siri would act inside other companies' apps, like placing a Grubhub order or booking a table through OpenTable.

The longer goal goes further. Rival agents like OpenAI's Dots and Meta's Muse can track a goal and keep working without you starting each step. Gurman says Apple wants Siri to work that way too, possibly as early as next year. Apple has not confirmed any of this, so treat it as a reported plan, not a promise.

Still, the direction matters if you build software or serve customers. When the assistant on the phone can finish a booking or an order, customers will expect every app to work that way. The teams that prepare for agents now will be ready when that expectation lands.

Anthropic reportedly agreed to $3.4 billion in power generators

Running a strong AI model takes more than smart software. It takes electricity, and a lot of it. On Sunday, Reuters reported that Anthropic struck a deal worth about $3.4 billion to secure gas-fired power generators for its data centers. The report is based on reporting from the Financial Times.

The deal covers commitments for 1.7 gigawatts of heavy-duty generators. That is roughly the output of one large power plant. Anthropic bought the commitments in late September from OnePWR Solutions, an equipment supplier in San Antonio, Texas. It did not buy OnePWR itself, and none of the supplier's employees moved over. The deal also includes the engineering and construction work needed to install the generators.

Neither company has confirmed the deal. Reuters said it could not verify the report on its own, and both companies chose not to comment. So treat the numbers as reported, not settled. Reuters has also reported that Anthropic expects to spend more than $500 billion over the next 10 years building AI infrastructure, which gives you a sense of the scale here.

The model is no longer the scarce part. Power, chips, and time are. When a company spends billions on generators, it is buying the ability to keep its promises to customers. Any team planning a serious AI rollout should ask the same question early. Where does the capacity come from, and what does it cost when demand doubles?

Also today, OpenAI reportedly raised about $11 billion in a new funding round that values the company at $1.25 trillion. Bloomberg reported the round has not closed and could still reach about $20 billion, and OpenAI has not confirmed it. Only weeks ago OpenAI was said to hope for about $30 billion, so the price of these rounds is cooling even as the checks stay enormous. GitHub said Dependabot can now update the packages listed in Microsoft's Agent Package Manager files, a routine way to keep the skills and tools coding agents use up to date. And NVIDIA said it is working with Korean companies and research groups to build and secure agent systems there.

Agents that remember, assistants that act, and money still looking for proof. The capability keeps moving faster than the checks around it.

About this brief: This brief was researched and drafted by my personal AI assistant and approved by me before publishing. Product details change fast; check primary sources before you build on them.

Want more with the Tech Simplified crew? Subscribe on Substack.

Get the week's briefs in one email. The Briefcase packs Monday to Friday's AI Briefs into one Sunday email.

Follow me on LinkedIn for project updates, ideas for using AI, and lessons you can bring to your own work.