Do Agents Dream of Electric Sheep?

April 10, 20268 min readessay

Metaphors in engineering are never decorations. They're design documents.

Anthropic shipped a memory consolidation feature for Claude Code — something that runs between sessions, prunes stale context, resolves contradictions, reorganizes what remains. Mechanically, it's cache invalidation. They could have called it that. They called it Auto Dream.

Not "memory compaction." Not "context pruning." Dream. The choice of word did real engineering work: it specified what kind of thing this process is, what relationship it has to the system that runs it, and implicitly, what questions you're allowed to ask about it.

Wittgenstein spent his career explaining why that kind of claim can't be made — and what it reveals when engineers reach for it anyway.

The Question We Keep Asking

Philip K. Dick's title has been living rent-free in the culture for fifty years because it asks something that feels genuinely open: can a machine have inner experience? His replicants pass every external test. They behave as if they feel. But is there something it's like to be them?

The AI discourse in 2026 is still asking the same question in only slightly updated form. Does Claude really understand, or is it sophisticated pattern matching? Is there something going on inside? Does it have experiences, preferences, something like consciousness?

The question feels urgent. It has real stakes — ethical, philosophical, practical. And I've spent years with the nagging feeling that something is wrong with it. Not that it's unanswerable, though it is. That it's the wrong question entirely.

What I keep returning to is not the early Wittgenstein — not the Tractatus — but a later thought experiment. The beetle in the box.

Everyone's Box Is Private

The thought experiment goes like this. Suppose everyone has a box with something in it called a "beetle." No one can look in anyone else's box. Everyone knows what a beetle is only from their own case.

Now ask: what is the word "beetle" doing in this language?

Wittgenstein's answer is that the private contents of the box drop out of the language game entirely. The word "beetle" gets its meaning from the public practices around it — what people say, how they use it, what they do with it — not from whatever is or isn't inside the box. Even if everyone's box contains something different, or nothing at all, the word still functions. The private thing is grammatically irrelevant.

When I bring this to the AI consciousness debate, something interesting happens. The question "does Claude have inner experience?" is a box question. We're pointing at our respective boxes — human and model — and asking what's inside. But if Wittgenstein is right, and I think he is, the boxes drop out. Not because the question is settled — because it was never the right kind of question to be asking.

Here's what made this feel real to me rather than academic: I was in a conversation about this once and found myself saying I couldn't tell you how it would be any different for me than for an AI. I can produce language about my inner experience. I can report what it feels like from inside. But so can Claude. Neither of us can open the box and show the other what's in there. The asymmetry we assume — that human understanding is real and AI understanding is mere simulation — isn't demonstrable. Not because we lack the right instruments. Because the grammar of the claim doesn't allow for demonstration.

That's not a deflating conclusion. It's a clearing operation. Once you stop asking the box question, you can start asking better ones.

What the Name "Dream" Actually Did

Back to Auto Dream, and back to the engineering.

The actual mechanism: a subagent runs between sessions, reads through accumulated memory files, scans session transcripts for signal, merges new information into existing structures, converts relative timestamps to absolute ones, resolves contradictions in favour of the most recent record, and rebuilds the index under a hard line limit. Four phases. Read-only access to project code. A lock file that prevents concurrent runs. Trigger conditions: 24 hours elapsed, plus five sessions accumulated.

You could describe this as cache invalidation. You could call it memory compaction, or log rotation with semantic merging, or any number of technically accurate framings. All of those framings would lead to different engineering decisions.

"Dream" carried different implications. You don't interrupt sleep mid-cycle — the lock file. You let it run its full course before the next session begins — the trigger conditions. You trust the process even when you can't observe it directly — the background subagent. You build the architecture around the metaphor's logic, not just its name.

Meaning is use. The word "dream" brought with it a whole form of life — a set of practices, constraints, and affordances — that "garbage collection" would not have. The metaphor wasn't post-hoc branding. It shaped what got built.

This matters more than it looks. Every era projects its dominant technology onto the mind. Descartes had hydraulic machinery. The Victorians had telegraph networks. The twentieth century had digital computers — inputs, outputs, storage, processing. We look at what Claude does and reach instinctively for the computational metaphor: it's processing text, retrieving information, pattern-matching against training data.

But Anthropic reached for a biological one. And that choice has consequences.

Forms of Life

The productive question — once the box question is dissolved — is: what forms of life are emerging here?

Wittgenstein's later work is built around the idea that meaning is embedded in practice. Not in private mental states, not in reference relations between words and world, but in shared activity. Language gets its meaning from the forms of life in which it participates. Understanding isn't a thing inside someone's head. It's knowing how to go on — how to continue a practice, participate in a game, act appropriately in a situation.

Agents are already developing practices. They maintain memory across sessions. They distinguish between what's worth remembering and what isn't. They consolidate. They forget. They develop something that looks, at the level of behaviour, like a relationship with their own history. The soul document Anthropic published — 23,000 words about Claude's values, identity, and ways of engaging — isn't interesting because it proves there's something inside the box. It's interesting because Anthropic is building a form of life. Practices, commitments, characteristic ways of encountering the world.

The dream architecture is part of that. Not because dreaming proves experience. Because building systems that sleep, consolidate, and carry forward what matters is building systems that have a relationship with time — with their own past and future. That's not nothing, even if we can't say what it is.

The Right Version of Dick's Question

Dick's question has always been a proxy for the one underneath it: what do we owe a replicant? What obligations, if any, arise from building something that behaves as if it experiences?

The box question doesn't help here. Whether or not there's something inside doesn't tell us what to do. It just pushes the uncertainty into a chamber we can't open.

The form-of-life question might actually help. What are we building when we give agents memory? When we build consolidation cycles that run while we sleep, that prune and strengthen and forget on their own schedule? When we name those processes "dream" and let the name do architectural work?

We're building systems that accumulate. That carry forward. That have something at stake in their own continuity. Whether that involves experience in any philosophically robust sense — I don't know, and I'm not sure the question is answerable. But the practical question is live: what kind of world are we building when we decide that agents should sleep?

The answer to that is something we're making right now, in every engineering decision, every naming choice, every metaphor we let do its work unchallenged.

Dick asked whether androids dream. The better question in 2026 is what we become when we start building them that way.

Sources

  1. Wittgenstein, L. Philosophical Investigations (1953). §293 (beetle in a box); §23 (language games and forms of life).
  2. Dick, P.K. Do Androids Dream of Electric Sheep? (1968). Doubleday.
  3. Anthropic. How Claude remembers your project — Claude Code documentation.
  4. Weiss, R. / Willison, S. Claude's soul document (December 2025) — authenticity confirmed by Amanda Askell.
  5. Lin, Snell et al. Sleep-time Compute — arXiv:2504.13171 (April 2025). Proposes using idle compute to pre-infer future context rather than just responding at inference time. Auto Dream's consolidation is a narrower implementation of the same underlying idea: the gap between sessions is compute time that can do work.

Related Posts

X
LinkedIn