AI Engineer Code Summit · Thu 12 Nov 2026 · Continental 4–6 (Mainstage) · Part 2 of 2

Understood backwards, lived forwards

An agent writes memory as it works, one session at a time, but memory is only useful if it makes sense as a whole. Anthropic's answer borrows from sleep: a separate pass, dreaming, where Claude reviews its transcripts and notes, fixes what's wrong, and builds better representations, such as a wiki, that load into the next session.

Lance Martin, Anthropic · last 11 min of an 18-min talk · Part 1: a thousand agents · Links and further reading
The problem

The paradox of agent memory

“So life can only be understood backwards, but it must be lived forwards.”00:07:06
“This is a fundamental paradox of agent memory because agents write memory typically as they go about a task.”00:07:06

An agent notices something important and writes it down for the task at hand. But memory has to be a coherent map later sessions can rely on:

“But memory needs to be something that's globally coherent and not redundant.”00:07:06

Humans have the same problem. During the day we form short-term memories; at night, it's believed, the hippocampus replays them to the neocortex, consolidating what matters into long-term knowledge.

“And we found agents can benefit from the same decomposition.”00:07:06
From the notes: “life can only be understood backwards, but it must be lived forwards”, the second time this week; the first was Michelle Chen's Cloudflare talk. Here it carries the argument: writing memory and making sense of it are two jobs, and they should run at different times.

Evidence: transcript 00:07:06

The evidence

Better models use memory better

On a benchmark of forty questions about a database, each in a fresh session, Claude can write to a memory directory in between. The only instruction:

“It just tells the agent the directory is available.”00:08:54

Older models barely benefit. Stronger ones gain much more from the same instruction:

“They know intrinsically when to write to memory, and they kind of know intrinsically when to read from memory to retrieve information.”00:08:54

Pokémon makes it visible. Claude 3.7's memory directory is enormous and indiscriminate:

“It creates this huge, unwieldy, just disgusting memory directory of things that are not really relevant or.”00:10:40

Fable 5's is small and strategic, and it gets much further in the game. On his slide: 413 files and 977,930 characters against 10 files and 43,406.

“So what we see is better models are much have much better instincts for salience.”00:10:40
From the notes: “Claude is getting better at reading and writing memory than it used to be.” The reading: the memory version of “Less harness, more model”: the memory system is just a folder; what improved is the model's judgment about what belongs in it.

Evidence: transcript 00:08:54, 00:10:40

The mechanism

Dreaming

Both examples are memory written during the work. Dreaming is a separate pass over everything stored.

“Transcripts you think about as kind of foundational.”00:12:13

A memory system could be just transcript search, but transcripts are large and most of what's in them doesn't matter. Memory notes are the next layer. Dreaming works on both:

“Dreaming lets you reflect over transcripts and these written notes, and re-represent them.”00:12:13

It can build new representations, such as wikis, and promote certain memories into context that loads into every future session.

From the notes: “I want to know what dreaming tokens are” and “follow up on building memory from session transcripts”. This answers both: dreaming is a compute pass that reads transcripts and notes and writes better memory, and dreaming tokens are what it spends. The reading: it automates what Steve Yegge does by hand when agent memory becomes a garbage dump, and it completes Lovable's design: keep the full log as evidence, render what reaches the model.

Evidence: transcript 00:12:13

Correction

Dreaming fixes poisoned memory

Playing Pokémon, Claude wrote down the wrong location for a landmark. From then on it mislocalized itself, fell through a trapdoor back to the ground floor, and repeated the loop.

“The key idea is a poison memory is injected and it's stuck there.”00:13:42
“Dreaming can rescue that.”00:13:42

Without dreaming, progress crashes at that point; with it, Claude corrects the note and carries on. The same pass fills gaps: a wiki Claude built had placeholder pages that didn't exist, and after dreaming they were filled and an index created.

“Dream is a way that Claude can reflect and kind of apply compute across its memory to improve it.”00:17:09

Evidence: transcript 00:13:42, 00:17:09

Representation

A wiki from a corpus

Anthropic acquired Stainless in May; from its documents and Slack messages, anonymized, they built 25 questions only that corpus can answer. Plain semantic search does reasonably well:

“You can see it's like 53 percent, just for search.”00:15:15

Then they let Claude build a wiki from the corpus.

“You can see as you push, you add more and more dreaming tokens.”00:15:15
“The wiki gets better and better and better, allowing Claude answers questions.”00:15:15
“We also find that adding the wiki and the search tool together actually bumps further beyond just raw.”00:15:15
From the notes: “wiki and data from memory”. The reading: for a client with a pile of documents, tickets and chat history, don't stop at search. Let a model spend compute building a structured map of the corpus, keep search alongside it, and rerun the consolidation as the corpus grows. It's a third answer beside Mixedbread's index everything ahead of time and Workday's judge every row: the model writes its own map of the knowledge. His post “The components of agent memory” is on the links page.

Evidence: transcript 00:15:15 · Back to part 1: a thousand agents