Anthropic concedes the new model knows nothing about your codebase, and CLAUDE.md is not the fix

At AI DevCon in London this summer, Lamis Mukta, a member of technical staff at Anthropic, told a room of developers that "the newest model we've just released isn't going to go out of the box and know exactly what it takes to succeed in your organization and what tasks you want it to do" ^[1][2]. She put the general case plainly: context is "often orthogonal to the model intelligence" at hand ^[3], which is a vendor saying out loud that buying a better model does not buy you situational knowledge.

The concession is specific. Mukta compared it to "agents [initially] not knowing their way around a codebase or knowing enough about your own user preferences" ^[4]. Her session, titled "Learning while you sleep, beyond memory to dreaming," then walked through what Anthropic has tried ^[1], and by her own account that is three architectures deep ^[11].

The first is the one most teams are running now: CLAUDE.md, a file Claude reads at the start of every conversation, carrying Bash commands, code style and workflow rules that the model cannot infer from the code itself ^[5]. Mukta's assessment was that it works to a degree but becomes hard to manage over time, particularly once a file holding important preferences gets very long ^[6]. Anyone who has watched a project file grow into an unreviewed policy document will recognise the shape.

The second is memory tools, where agents decide autonomously when to read, write and update memories ^[7]. That happens in-band, inside the session ^[8]. Paired with progressive disclosure, the agent scans light metadata at layer one and only pulls full content and original source files at layers two and three, so the context window does not get flooded ^[9]. Mukta's analogy was a bookshelf: scan the titles, pull the one book that looks relevant ^[10]. The limits she named are the interesting part. Humans and agents are still jointly deciding what deserves to become a skill ^[12], memories go stale ^[13], and a memory file can be written wrongly or maliciously injected, which is why guardrails are needed ^[14].

The third is "dreaming," an asynchronous batch process with its own allocated resources that consolidates memory, cuts what is no longer relevant, adds what agents are missing, and reorganises the store ^[15][16]. Being out-of-band is the point: agents doing this in-band must split effort between executing tasks and curating memory for their future selves ^[17]. Dreaming looks for recurring failure patterns, including wrong units, missing topics, broken tool configs and stylistic tics such as overused em dashes, and proposes updates for human review ^[18]. Anthropic says this architecture underpins its Managed Agents memory and API approach ^[19].

The obvious objection came from outside. Jayakumar Ramalingam, a staff software engineer, cloud architect and independent researcher in AI agent systems, told The New Stack that "dreaming is useful, but it also creates a dangerous promotion path" from repeated mistakes to persistent policy ^[20]. His framing: "A bad answer normally dies with the session" ^[21], while a bad memory does not.

What to watch is the review step. Dreaming proposes changes for human approval ^[18], but the reported presentation does not describe how those proposals are surfaced, batched or rejected at volume ^[22]. A consolidation process that nobody has time to audit is just a longer CLAUDE.md with a scheduler.