In July 2026, Anthropic's interpretability team published Verbalizable Representations Form a Global Workspace in Language Models - the first mechanistic account of how a language model holds thoughts in mind. Their instrument, the Jacobian lens, reads out which concepts a model is poised to put into words at any moment of its processing. What it found was a two-tier mind.
Most of a model's activity is automatic. Parsing, grammar, routine recall - it all runs beneath the surface, inaccessible even to the model itself. Above that sits a small, privileged workspace: a few dozen verbalizable concepts, holding the intermediate steps of whatever the model is deliberately reasoning about. Ask a model how many legs the animal that spins webs has, and spider sits in that workspace mid-computation - unspoken, but causally load-bearing. Swap it out, and the answer changes.
Three properties of that workspace matter for anyone building memory for agents.
It thinks in words. The workspace isn't an abstract vector soup - it's organized around representations of things the model might say. Reasoning routes through verbalizable concepts. Context that arrives as natural prose is already in the format the substrate computes with; context that arrives as structured payloads has to be translated first.
It is small and contested. The workspace holds on the order of tens of concepts, and new content evicts old. Injecting ten marginally relevant memories isn't harmless filler - it competes for the same slots as the things that matter. Retrieval precision beats retrieval volume.
It reads everything from the assistant's point of view. Post-training installs a perspective: the workspace forms reactions - safety assessments, empathy, objections - while the model is still reading the user's message. Memory framed as the agent's own recall lands in machinery that is already listening from that vantage point.
adja's design lines up with all three. Memories are stored and recalled as natural language, not payloads. Salience ranking and token budgets keep recall sparse and precise. And memory is framed in the first person - accumulated knowledge, not database lookups. We didn't design adja from this paper; it didn't exist yet. We designed it from a conviction that an agent's memory should work the way its mind does. The research now shows what that way is.
One more finding worth sitting with. The paper shows that models extend their limited internal workspace by externalizing - writing intermediate thoughts out as text and reading them back. An agent's context window is its externalized workspace. The question of what belongs in it, at every session start, is the question adja exists to answer.