How prior conversation is maintained when reading an utterance

The question

The same phrase, “I think that works,” means different things depending on what has been said before. Reading the meaning and connections of a thought requires knowing the context in which the utterance was made.

This page addresses the question: when reading an utterance, how much of the prior conversation is retained, and in what form?

Principle: subsequent utterances are not read

At the time of speaking, speakers do not know how the conversation will unfold. Similarly, when AWAI reads an utterance, it treats only the preceding conversation as context. Thoughts that appear later are never added to the context.

AWAI is deliberately designed not to use information that could only be known by looking across the entire conversation, in order to read what the utterance meant at that moment.

The flow shared in the room

Participants move a conversation forward by building up what has been said as shared understanding. Clark and Brennan (1991) call this “common ground” and describe participants updating it moment by moment as the conversation proceeds. Stalnaker (2002) likewise treats common ground as information shared among the participants that changes over the course of a conversation.

AWAI maintains this “flow shared in the room” in two forms:

• the immediately preceding utterance, as the thoughts extracted from it
• before that, a brief summary of each utterance

The literature on common ground, however, does not specify how much information people can hold at once. AWAI therefore draws its guidelines for how much to retain from research on working memory: that people can hold only about three to five chunks at a time (Cowan 2001), and that a system exists to hold information grouped by meaning, bound to long-term memory (Baddeley 2000). What is retained as context follows the idea of common ground; how much is retained is based on what the human mind can hold.

Any other thoughts within the utterance being read are also included as context.

Association

In conversation, a topic raised much earlier can suddenly connect with what is being said now. Using the current utterance as a cue, AWAI recalls past thoughts that are close in meaning, even when they occurred much earlier in the conversation.

A neurocognitive model of schemas from clinical psychology (Moscovitch et al. 2023) likewise describes memories as fading over time while being recalled by the present situation and its cues. AWAI's “association” similarly retrieves past thoughts based on how close in meaning they are to the current utterance; the further back a thought is, the closer in meaning it has to be.

Preconceptions (schemas)

People interpret new events through frameworks formed by past experience. Psychology and neuroscience call these “schemas.” A schema makes related memories easier to recall and shapes how new information is received (Gilboa & Marlatte 2017; Bein & Niv 2025). The neurocognitive model mentioned above (Moscovitch et al. 2023) also holds that memories congruent with an active schema are recalled more easily.

AWAI maintains these “preconceptions (schemas)” for each speaker.

• BuildAmong the thoughts a person has voiced in this conversation, dense clusters of thoughts close in meaning are gathered as that person's preconceptions (HDBSCAN, a density-based clustering method: Campello et al. 2013, building on DBSCAN, Ester et al. 1996).
• KeepAs the conversation proceeds and a cluster changes, its summary is updated and the previous version is retained.
• RecallWhen the person makes a new utterance, clusters related to it are recalled and passed as premises for reading that utterance. The LLM also judges whether a recalled preconception supports or opposes what is currently being said, and incorporates that into its reading.

The only material used is what the person has said within this conversation; experience outside the conversation is not included.

Recurring thoughts carry more weight

A thought whose content has recurred repeatedly is easily recalled by that person.

The cognitive architecture ACT-R computes how easily a memory can be recalled (its base-level activation) from when and how often it has been used (Anderson et al. 2004). The equation rests on the analysis by Anderson and Schooler (1991) that the pattern of past occurrences predicts the probability that a memory will be needed again.

AWAI applies this by reinterpreting “how often it has been used” as “how often similar content has appeared in the conversation.” Among the thoughts recalled through association, those that belong to a recalled preconception and have a sufficiently high ease of recall are treated as especially important context.

Context not currently handled

The only context AWAI can use is the words spoken in the conversation. Information such as experiences outside the conversation, assumptions taken for granted by participants, tone of voice, facial expressions, and the atmosphere in the room cannot be handled.

A person's preconceptions are formed over an entire lifetime. AWAI's preconceptions, by contrast, are constructed within a single conversation and remain very limited. If this kind of information could also be taken into account, we believe the reading would become more accurate.

References for this page

• Clark, H. H. and Brennan, S. E. (1991). Grounding in communication. In L. B. Resnick, J. M. Levine and S. D. Teasley (eds.), Perspectives on Socially Shared Cognition, pp. 127–149. American Psychological Association. doi.org/10.1037/10096-006Note — what this page draws on:The view that participants build what has been said into common ground and update it moment by moment as the conversation proceeds (grounding). It underlies the idea of keeping the flow shared in the room.
• Stalnaker, R. (2002). Common ground. Linguistics and Philosophy, 25(5–6), 701–721. doi.org/10.1023/A:1020867916902Note — what this page draws on:The definition of common ground as information shared among the participants that changes over the course of a conversation.
• Cowan, N. (2001). The magical number 4 in short-term memory: A reconsideration of mental storage capacity. Behavioral and Brain Sciences, 24(1), 87–114. doi.org/10.1017/S0140525X01003922Note — what this page draws on:The capacity of short-term memory: about three to five chunks at a time. It serves as a guideline for how much information the flow retains.
• Baddeley, A. (2000). The episodic buffer: a new component of working memory? Trends in Cognitive Sciences, 4(11), 417–423. doi.org/10.1016/S1364-6613(00)01538-2Note — what this page draws on:The episodic buffer, which holds information grouped by meaning as one unit bound to long-term memory. It grounds keeping the flow as per-utterance summaries.
• Gilboa, A. and Marlatte, H. (2017). Neurobiology of Schemas and Schema-Mediated Memory. Trends in Cognitive Sciences, 21(8), 618–631. doi.org/10.1016/j.tics.2017.04.013Note — what this page draws on:Describes a schema as a framework formed from past experience that makes related memories easier to recall and shapes how new information is received. It underlies treating schemas as preconceptions.
• Bein, O. and Niv, Y. (2025). Schemas, reinforcement learning and the medial prefrontal cortex. Nature Reviews Neuroscience, 26, 141–157. doi.org/10.1038/s41583-024-00893-zNote — what this page draws on:Shows that schemas guide memory retrieval and that events sharing common features are processed together. It grounds bundling clusters of similar thoughts into each speaker's preconceptions.
• Moscovitch, D. A., Moscovitch, M. and Sheldon, S. (2023). Neurocognitive Model of Schema-Congruent and -Incongruent Learning in Clinical Disorders: Application to Social Anxiety and Beyond. Perspectives on Psychological Science, 18(6), 1412–1435.Note — what this page draws on:Covers the general properties of schemas from this clinical neurocognitive model: that memories fade over time yet are recalled by the present situation and its cues, and that memories congruent with an active schema are recalled more easily.
• Ester, M., Kriegel, H.-P., Sander, J. and Xu, X. (1996). A Density-Based Algorithm for Discovering Clusters in Large Spatial Databases with Noise. Proceedings of KDD-96, pp. 226–231.Note — what this page draws on:The principle of treating data points connected by density as a single cluster (DBSCAN), serving as the foundation for building preconception clusters.
• Campello, R. J. G. B., Moulavi, D. and Sander, J. (2013). Density-Based Clustering Based on Hierarchical Density Estimates. PAKDD 2013, LNCS 7819, pp. 160–172. doi.org/10.1007/978-3-642-37456-2_14Note — what this page draws on:A method that extracts stable clusters without requiring a fixed density threshold (HDBSCAN). It is the method AWAI uses to build preconception clusters.
• Anderson, J. R., Bothell, D., Byrne, M. D., Douglass, S., Lebiere, C. and Qin, Y. (2004). An Integrated Theory of the Mind. Psychological Review, 111(4), 1036–1060. doi.org/10.1037/0033-295X.111.4.1036Note — what this page draws on:ACT-R's base-level learning: the equation that computes how easily a memory can be recalled from when and how often it has been used. It serves as the basis for reinterpreting usage frequency as how often similar content has appeared in the conversation.
• Anderson, J. R. and Schooler, L. J. (1991). Reflections of the Environment in Memory. Psychological Science, 2(6), 396–408. doi.org/10.1111/j.1467-9280.1991.tb00174.xNote — what this page draws on:The analysis that the pattern of past occurrences predicts the probability that a memory will be needed again, which grounds the equation above.