All my life, I've occasionally woken up with a new idea or a solution to a nagging problem. I even have a waterproof pad and pencil in my shower to record the thoughts. I call it dream solving. So I was taken aback when Anthropic announced Dreams to improve Claude's memory.
Every Claude user hits the same wall. You explain your business on Monday, and on Tuesday Claude has forgotten it. You paste the same background into a new chat, re-explain who your customers are, and correct the same mistake you corrected last week. The model is brilliant, and it starts every session as a stranger.
On November 6, Anthropic published the most detailed account I have seen of how it thinks about that problem, a research post called "Components of agent memory" [1]. It sits next to a feature Anthropic has been testing with developers since this spring, called Dreams, which lets a Claude agent go back over its past work between sessions and rewrite its own memory [2]. I read the article, the docs and the coverage. This post covers what Anthropic found, what Dreams does and does not do, and how it lines up with Start, the file-based company memory we have been running our own agents on since early October and now ship with Chet.
The short version: Anthropic's research describes memory as a stack of layers, and it lands close to where we landed after four major versions and at least 32 smaller revisions. The part that is new to me is the idea of a scheduled, offline pass that cleans memory up. We are going to borrow it.
What Anthropic means by agent memory
Anthropic defines agent memory narrowly: knowledge Claude gains from experience and keeps between sessions [1]. It explicitly leaves out three things people often lump in. The model's weights are not memory in this sense. Neither is what Claude picks up inside a single conversation, and neither is compaction, the trick of summarizing a long conversation so it fits [1]. Memory is what survives after the session ends.
The article then breaks memory into parts. I would group them as four layers.
An event log. The raw record of what happened: session transcripts, searchable after the fact. Anthropic's conclusion near the end is that searching this log may be the foundation everything else rests on [1].
Notes. Things Claude decides are worth writing down, either by tagging moments in the log or by saving notes to a memory folder [1]. One finding here surprised me. When Anthropic told Opus 5 to "record only failures," it lost much of the boost memory had given it. Letting the model judge what mattered worked better [1]. Salience is a skill, and newer models have more of it.
Representations. Notes get messy, so Anthropic had Claude maintain wikis: linked pages it creates, merges and prunes over time [1]. On a test built from one company's documentation, a Claude-maintained wiki went from meeting 13% of the graded criteria after 2 sessions to 48% after 128 sessions. A wiki plus search scored 61%, against 53% for search alone [1].
Injected context. The small top layer that loads at the start of every session, before Claude does anything. Anthropic is blunt about the cost: every injected line spends tokens, and, as its own Claude Code documentation warns, longer instruction files reduce how well Claude follows them [1][3]. Claude Code puts a hard number on it. It loads only the first 200 lines or 25 KB of its memory index file, whichever comes first [3].
How much does it matter? Anthropic built a benchmark where Claude answers a series of database questions, each in a fresh session, with only a memory folder carried between them. Opus 5.5 scored 0.66 without memory and 0.84 with it, on a reward scale of 0 to 1 (a score, not a percentage), and other recent models made similar jumps [1]. That chart is the strongest argument in the article that memory is worth the trouble.
Two smaller points matter as much as the layers. First, Anthropic found it works better to tell Claude what memory is for than to script how to use it [1]. Second, models need to be told the memory exists. Sonnet 4.6 checked its memory folder before answering less than half the time; telling it to read memory first nearly doubled its score [1]. A memory nobody reads is a filing cabinet in a locked room.
What Dreams does
Dreams is the piece that got the headlines. The name comes from an analogy to sleep: people consolidate the day's experience overnight, and Anthropic wanted an agent that does the same between sessions [1].
Here is how it works, from the documentation [2]. A developer using Claude Managed Agents, Anthropic's hosted agent service, calls the Dreams API with one existing memory store and up to 100 past session transcripts. Claude reads them, checks the notes against what happened, fixes errors, merges duplicates, and writes new notes and wiki pages. In one of Anthropic's examples, a dreaming pass caught a wrong note about the Pokémon game, checked it against the transcript, and rewrote it [1]. A run takes from minutes to a few hours and is billed at normal API token rates for whichever model you choose [2].
The design choice I like most: Dreams never edits the original memory. It writes a new memory store, and a person decides whether to attach it to future sessions or throw it away [2]. The agent proposes, a human approves.
Now the limits, because they matter for anyone deciding what to do this week. Dreams is a research preview. It is available through the Claude API to developers who request Managed Agents access, and it needs two separate beta flags to switch on [2]. It is not in the Claude apps, and it is not part of Pro, Max, Team or Enterprise plans as a feature you can turn on. Nothing runs on a schedule unless a developer builds the schedule. It is not eligible for zero data retention or HIPAA [2]. Some blogs report a related "Auto Dream" in Claude Code that consolidates its memory files automatically; I could not find it in Anthropic's own Claude Code documentation, so I would treat it as unconfirmed for now [4][3].
In plain terms, Dreams is a developer building block today, not something a founder running a business in Claude can use on Monday.
Someone already built the consumer version
Eleven days after Anthropic's article, Ben van Sprundel of Ben AI, a popular AI educator, released a free /dream skill for Claude Code and Cowork and explained it in a video [5]. It rereads your recent chats every night, fixes notes that turned out to be wrong, adds new lessons, merges duplicates and shows every proposed change on one page for you to keep or revert. He splits the work across a lead agent, several analyzer agents and one organizer that alone may rewrite the memory index. That split is his design, not Anthropic's, and it is a sensible one.
Two things stand out. First, demand is real. People can feel their Claude memory going stale and want something to fix it, and they will take the fix from whoever ships it first. Second, his skill dreams over one person's memory. He says it himself: the lessons one person learns in their own chats normally never reach the rest of the team. His skill can propose a rule for a shared team file, but the memory it maintains is still yours alone.
That gap is the whole reason Start exists. A company does not need each employee's Claude to get smarter in private. It needs what any one of them learns to land in a memory everyone's agents read.
Where Start lines up, layer by layer
When I read Anthropic's list of components, I kept recognizing folders. That is not because we were clever. It is because anyone who runs agents on real work for a few months runs into the same failures and ends up with the same fixes.
The event log is our events folder. Every agent writes an append-only day file: one line per thing that happened, with a time, a type and a pointer to the detail. Lines are never edited; corrections are new lines. Anthropic's view that a searchable log is the foundation matches ours. When an agent is unsure what happened, the log settles it.
Notes are our lessons files. Each agent keeps a short file of lessons learned, and we let the agent decide what goes in. Anthropic's finding on salience, that rigid "record only X" rules hurt stronger models, is a reason to keep it that way.
Representations are our dossiers. One file per subject, a customer, a product, a tool, each with dates, sources and a history section where changed facts are struck through rather than deleted. That is a wiki maintained by Claude, with one extra rule we learned the hard way: every fact carries the date it was checked.
Injected context is our AGENTS.md. One file at the top of the folder that loads every session and points to everything else. Ours has a hard cap of 40 lines. We added the cap after an earlier version of our setup gave one agent a 140 KB task file with 174 open items, and the agent worked only on whatever was near the top. Anthropic's warning that longer instruction files reduce adherence [1] is the same lesson with better data behind it.
Steering is our "when" lines. Anthropic recommends telling Claude what memory is for. Every file in Start opens with a line that says exactly when to read it, such as "before writing anything the public will see". The agent learns the purpose of each file before it reads the contents.
Where we differ is ownership. Anthropic's memory belongs to one agent, inside one hosted service, reached through an API. Start belongs to the company. It is a folder of plain markdown in the company's own Google Drive, readable by a person in any text editor, shared by every agent the company runs, and usable from Claude's apps, Cowork or Claude Code. If you switch hosts, or a new Claude feature changes how projects work, the memory stays put.
What we are changing because of this
Dreams made one gap in our setup obvious. Start has a roll-up procedure that summarizes old event logs by month and quarter, but nothing goes back over recent work and asks, "Which notes are now wrong? Which dossier is stale? Which lesson has been learned three times by three different agents?" Today a person catches those, usually me, usually late.
So we built one. Since November 18, Chet runs a nightly dream pass on our own Start folder, a company-level version of what the consumer skills do for one person. It reads the last week of event logs and the files they touched, then proposes specific edits: a fact to correct, two notes to merge, a lesson to promote into a rule. The first run read 593 log lines written by 10 agents and proposed seven changes. Four agents had learned the same lesson about verifying work in one week, so it proposed making that a company rule. It also caught a file still quoting a price we had since changed twice. Following Anthropic's pattern, nothing changes until a person approves it. It does not need the Dreams API, because the memory is already in files Claude can read in any host. Chet members get it next.
We are also tightening the top layer. If longer instruction files reduce adherence, every line in AGENTS.md has to earn its place, and Chet's health check should flag any setup where the always-loaded files have grown past what Claude will reliably follow.
Current view (subject to change)
I think Anthropic is right that memory is a hierarchy, and right that the hard part is maintenance, not storage. Writing things down is easy. Keeping them true is the job.
I also think per-agent memory inside a lab's hosted service will be the default for developers, and that is fine. Most companies do not run one agent. They run Claude in a few places, used by a few people, and need all of it to share one memory they can read, audit and take with them. That is the problem we built Start to solve.
What would change my mind: if Anthropic ships Dreams-style consolidation into the Claude apps with company-level, human-readable memory that works across Projects and Cowork, a separate memory folder becomes much less necessary. I would welcome it. Until then, the method in Anthropic's research is available to anyone who keeps their memory in files.
Final Thoughts
The most useful thing about Anthropic's article is that it is plain about cost. Injected memory costs tokens. Dreaming costs compute. Long instruction files cost adherence. Every layer is a trade, and Anthropic's own tests show the gains come from combining layers, not from any one of them: a wiki plus search beat search alone by eight points [1].
If you use Claude for your business and want to act on this today, three steps cover most of it. Keep one short file Claude reads at the start of every session, and cap it. Keep a dated log of what happened, separate from your notes. And once a week, have Claude reread the log and propose corrections to your notes, which you approve. That is dreaming by hand, and it works in the Claude you already pay for.
If you would rather not build it yourself, that is what Chet and Start are for. Chet sets up the folder in your own Google Drive, grades how well your setup uses Claude, and keeps it current as Claude changes. Maximize Claude with Chet, $99 a month at coworkers.global.
Regards,
Charles Stack
Founder, Coworkers.Global
Coworkers.Global is not affiliated with Anthropic.
Sources
[1] Anthropic, "Components of agent memory," published November 6, 2026. https://anthropic.com/research/components-of-agent-memory
[2] Claude Platform docs, "Dreams" (Claude Managed Agents, research preview). https://platform.claude.com/docs/en/managed-agents/dreams
[3] Claude Code docs, "Manage Claude's memory." https://code.claude.com/docs/en/memory
[4] ClaudeFast, "Auto Dream" (third-party, unconfirmed by Anthropic). https://claudefa.st/blog/guide/mechanics/auto-dream
[5] Ben AI, "Anthropic Just Revealed How To 10x Your Claude Memory," YouTube, November 17, 2026. https://www.youtube.com/watch?v=uYdbxzxRWl8