The language we use to describe AI shapes how we understand what it can and should do—and right now, we're setting ourselves up for a confusion crisis.
The Summary
- Anthropic announced new features called "dreaming" and "memories" for AI agents at its developer conference, adding to a growing trend of anthropomorphic AI feature naming
- The metaphor gap between what AI actually does (pattern matching, data processing) and what these names suggest (consciousness, subjective experience) isn't just marketing fluff—it's creating fundamental misunderstandings about capability and risk
- When users expect "thinking" but get statistical prediction, the disconnect doesn't just disappoint—it breaks trust and makes it harder to build reliable agent systems
The Signal
Anthropic's choice to call its new agent features "dreaming" and "memories" isn't an isolated branding quirk. It's the latest data point in a pattern that spans the entire AI industry. OpenAI has "reasoning." Google has models that "understand." Startups pitch agents that "learn" and "remember." Every feature release sounds less like software and more like raising a digital child.
The actual mechanics are far more mundane. What Anthropic calls "dreaming" is likely some form of background processing where agents consolidate interaction data. "Memories" are probably just better context windows or vector databases that let models reference previous conversations. These are legitimately useful capabilities. They're also nothing like their biological namesakes.
"The metaphor gap isn't just annoying—it's actively making it harder to deploy agents reliably."
This matters because people build mental models from language. Call something "memory" and users expect it to work like human memory: persistent, retrievable, connected to identity. When it doesn't—when the agent forgets yesterday's conversation or hallucinates a detail—the failure feels personal. The agent lied. The agent broke its promise. Except it didn't, because it never had those capabilities to begin with.
The naming problem cascades into deployment decisions. If your sales team believes the agent "understands" customer intent because that's how the vendor described it, they'll hand it conversations it can't actually handle. If leadership thinks agents "learn" from every interaction, they won't build the feedback loops and human review systems that actually make agents better over time.
Here's what these features actually do:
- "Dreaming" = batch processing of accumulated data during low-activity periods to update model weights or embeddings
- "Memory" = maintaining conversation context across sessions via databases, not encoding experiences into model parameters
- "Reasoning" = chain-of-thought prompting that makes intermediate steps visible, not symbolic logic or causal understanding
The gap between the poetry and the plumbing creates two concrete problems for anyone building with these tools. First, evaluation frameworks break down. How do you measure whether an agent "dreams" effectively? You can't, because the metaphor doesn't map to testable metrics. You end up with vague success criteria and projects that drift.
Second, the language makes it nearly impossible to explain failures to stakeholders. When an agent with "memory" forgets a key detail, explaining that it's actually a context window limitation or a retrieval ranking issue sounds like excuse-making. The name promised human-like capability. The excuse reveals it's just software with edge cases.
The Implication
If you're building with AI agents, insist on technical precision from your vendors. Ask what "memory" actually means in architectural terms. Request documentation that describes capabilities without metaphor. Build your own internal vocabulary that maps to how these systems actually work: "context persistence," "background indexing," "multi-turn state management."
For AI companies: the cute names might win you clicks, but they're poisoning the well. Every overpromise makes the next deployment harder. The companies that win Web4 won't be the ones with the cleverest metaphors—they'll be the ones whose customers understand exactly what they bought and can deploy it reliably at scale.