How do humans maintain separate mental contexts during a single conversation?
This explores how people hold multiple parallel threads — separate beliefs, intentions, and topics — in mind at once during a single conversation, and what that machinery reveals by contrast with how LLMs handle the same thing.
This reads the question as being about the cognitive machinery humans use to keep several mental threads alive at once inside one conversation — separate beliefs, intentions, and topic stacks — rather than memory across sessions. The corpus suggests humans don't do this as a single trick but by running several layers in parallel. The clearest statement is that discourse comprehension demands tracking three irreducible layers simultaneously — the literal segments being said, the intentional structure (why each segment is uttered), and attentional salience (what's currently in focus) — and that these constrain each other rather than running in sequence How do readers track segments, purposes, and salience together?. Your 'separate mental contexts' are largely that attentional layer: a focus stack that lets you push a digression and pop back to where you were.
A second piece of the machinery is that humans maintain a model of the *other* person's context, not just their own. Communicative grounding is person-specific — the same words point at different things for different people — so you're constantly negotiating which meaning is shared Why do speakers need to actively calibrate shared reference?. Formal work on this treats dialogue as bidirectional belief tracking, where each speaker carries and updates a model of both sides' beliefs as turns progress from partial to shared understanding Can dialogue systems track both speakers' beliefs across turns?. So one of the 'separate contexts' you keep is literally a context belonging to someone else's head.
What actually holds the threads together between detours is invisible relational work. Humans keep a conversation coherent through implicit maintenance techniques — reference repair, topic hand-off, marking when you're stepping aside and when you're returning — that aren't about transmitting information at all but about sustaining the shared frame Why don't language models develop conversation maintenance skills?. These are the seams that let you juggle contexts without the conversation falling apart, and they leave detectable fingerprints: speakers unconsciously entrain on each other's vocabulary and style, and that coordination shifts measurably under load — it rises, for instance, during deception Do liars and listeners coordinate their language during deception?.
The sharpest insight comes from the contrast with LLMs, which the corpus circles repeatedly. A human prompt to a model collapses utterance, context-assignment, and role into one static frame the model can't renegotiate mid-stream — so where you fluidly pivot and re-focus, a model needs explicit re-prompting How do prompts reshape the role of context in AI conversation?. Models also don't develop the maintenance skills above, because training rewards predicting information, not relational upkeep Why don't language models develop conversation maintenance skills?, and they lack lexical entrainment entirely Why don't conversational AI systems mirror their users' word choices?. Underlying all of it is a structural asymmetry: humans have a continuous biological substrate that carries interaction effects even through silence, while a model instance is reconstituted from stored text each time Does an LLM have anything that persists between conversations?.
The thing worth taking away: maintaining separate contexts isn't mainly a feat of storage. It's a feat of *coordination* — running a focus stack, a model of the other person's beliefs, and a layer of implicit repair work all at once — and it's precisely the coordination layer, not the memory, that current AI systems are missing.
Sources 8 notes
Discourse processing demands parallel recognition of linguistic segments, intentional structure, and attentional salience—not sequential processing. These three layers constrain each other during comprehension, and failures in any single layer disrupt overall understanding.
The same words can mean different things to different speakers because referential grounding is person-specific. True communicative grounding demands collaborative negotiation of how language connects to the world, not mere surface-level word sharing.
CRSA integrates rate-distortion theory with RSA to enable bidirectional belief tracking across dialogue turns. Demonstrated on referential games and doctor-patient dialogues, it captures progression from partial to shared understanding, providing the information-theoretic framework that token-level LLM systems lack.
Humans keep conversations smooth through implicit techniques like reference repair and topic hand-off that sustain relational interaction, not convey information. Language models don't develop these because training signals reward information prediction, not relational work.
Research shows interlocutors' linguistic styles correlate more during false communication than truthful communication, especially when the speaker is motivated to deceive. This coordination serves as a detectable deception signal through the listener's adaptive behavior, not just the liar's language.
LLM prompts bundle utterance, context assignment, and role specification into a single static frame the model cannot renegotiate, unlike human dialogue where context evolves cooperatively. This makes mid-conversation pivots require explicit re-prompting rather than implicit adjustment.
Response generation models fail to adapt vocabulary toward users' lexical choices, a phenomenon central to human rapport and clarity. Post-training via DPO on coreference-identified preferences can teach models in-context convention formation.
While humans have a continuous biological-phenomenological substrate that preserves interaction effects during dormancy, LLMs have no analogous carrier. The virtual instance is reconstituted from stored text each time, making resumed and new conversations structurally identical.