Anthropic’s latest engineering report explores how Super Intelligence (SI) agents can turn past work into better notes—and put those lessons to use on the next task.
American Super Intelligence IndexPublished · News & analysis
A better handoff.
A useful assistant should not need the same lesson twice. If yesterday’s work revealed how a project is organized or which approach actually solved a problem, tomorrow’s task has a better starting point when that knowledge survives.
In a November 6 engineering report, Anthropic describes experiments with a practical way to build that continuity: give Claude saved notes, then help it decide what belongs in them, check them against past work and organize them for later use.
The constructive advance is not an infinitely expanding conversation. It is a more useful notebook. An agent can start a fresh session while retaining selected information in files it can consult when needed.
Keep the lesson, not every detail.
Anthropic distinguishes this kind of memory from retraining a model or summarizing a conversation to fit a shorter context window. Here, the durable information lives outside the model, in a store that later sessions can read.
The company’s experiments suggest that useful notes capture more than a running diary. Tested rules, the evidence behind them and a workable plan can matter more than recording every passing detail. Facts already available from the current environment may not need a second, increasingly stale copy.
That makes memory an editorial task as well as a storage task: preserve what is likely to help, and retain enough context to understand why it was saved.
Check the notebook between jobs.
The report also explores “dreaming,” an engineering nickname for processing past records between tasks. There is no claim of sleep or consciousness. The system reviews notes alongside session transcripts, giving it an opportunity to correct a mistaken entry or combine related information.
In one game-based experiment, checking a wrong location note against the original transcript helped Claude recover its route. In a separate document-search evaluation, Anthropic found that linked, wiki-style notes could complement a search tool. These are specific company-run experiments, not proof that every task improves automatically.
The Claude Managed Agents documentation describes Dreams as a research-preview feature. It produces a separate, reorganized memory store without modifying the input store, so a developer can inspect the result before using it in future sessions. This report is not an announcement that every Claude conversation now has that behavior.
More continuity, less repetition.
For a workplace assistant, the analogous benefit could be simple: fewer repeated explanations about a codebase, better recall of a verified project decision and a clearer handoff when work resumes. Those are practical possibilities, not outcomes established for every workplace by this research.
Our reading: useful progress in SI also comes from improving the support systems around a model. A smaller set of well-supported notes, checked against the record and easy to navigate, could make a capable assistant more consistent over a long project. Sometimes the next step forward is a better way to remember the last one.
