ETD

• affective computing
• episodic memory
• interactive conversational agents
• large language models
• long-term memory
• semantic knowledge graph
• stigmergy

Long-term interaction between a person and an Interactive Conversational Agent requires the agent to maintain a persistent, evolving model of that person across many separate conversations. Current architectures rely on a short context window and have no mechanism for retaining, consolidating and retrieving what they learn.
When a person talks to an assistant over weeks or months, they expect it to know them. They expect not to explain again who their sister is, not to be asked the same question twice, and not to be told something that contradicts what they said last month. Conversational assistants today cannot do this. Each conversation begins with nothing, and everything is gone the moment the session closes. The obvious remedy is to keep the transcript and read it back later. But a transcript grows without limit and almost all of it is irrelevant to any particular question.
This thesis takes a different approach, borrowed from the way human memory is described in cognitive psychology. It is a cognitively inspired memory architecture that addresses those limitations, together with its implementation and an empirical evaluation. The system never stores the conversation, it stores what it understood.
The system holds memory at three timescales. A volatile layer keeps the current conversation, the tracked mood of the user, and a short-term planning buffer carrying the agent’s own pending commitments. An episodic memory stores self-contained summaries of what happened. A semantic memory holds a knowledge graph of the people, places, objects and topics in the user’s life, with the relations between them. Every long-term memory is text the system generated itself, and raw turns are discarded once they have been summarised.
Both memory types carry affect. Each episode records the emotions attributed to it, and each entity in the graph carries a stigmergic trace: a sparse field over the valence–arousal plane onto which every episode mentioning that entity deposits, and which decays when the entity is neglected, so what the agent knows about a person includes not only what is true of them, but what has been difficult, or good, or worrying.
Retrieval ranks episodic and semantic candidates through a composite score whose components (semantic similarity, recency, affective intensity, congruence with the user’s current mood and the number of supporting episodes) are bounded on a common scale, so that the two memory types remain directly comparable. Seman tic retrieval also exploits a degree-corrected personalized PageRank over the user’s graph, which recovers memories that are relevant by connection.
The thesis presents the design of this architecture, its implementation as a work ing system, and its evaluation on public datasets of long conversations such as Lo CoMo and LongMemEval and of emotional narratives with IDEST. It reports what the evaluation established, what it ruled out, and where the architecture’s remaining limits lie.