Hello Again! LLM-powered Personalized Agent for Long-term Dialogue

Modern conversational artificial intelligence systems frequently struggle to sustain long-term engagement across multiple interaction sessions. While large language models excel at brief, single-session exchanges, practical applications—such as personalized companions, virtual assistants, and customer support—demand that agents remember past events, track conversational history over days or years, and preserve consistent personal traits. Most existing systems either isolate event memory from user profiling or rely heavily on rigid architectures that fail to generalize across diverse operational domains.

The article introduces and evaluates the Long-term Dialogue Agent (LD-Agent), a modular, model-agnostic framework designed to deliver coherent, personalized interactions across extended multi-session dialogues. The primary objective is to demonstrate that coordinating dynamic event perception with bidirectional persona modeling significantly enhances conversational quality, character consistency, and domain adaptability.

To achieve this, the authors designed a three-part modular system comprising an event memory perception unit, a persona extraction unit, and a response generation unit. The event memory perception unit separates historical session summaries in a long-term memory bank from ongoing dialogue context in a short-term memory cache, employing a retrieval mechanism based on semantic relevance, topic noun overlap, and time decay. The persona extraction unit dynamically extracts and updates personality traits for both the user and the agent using fine-tuned instruction models. The authors evaluated the framework on standard five-session benchmarks (MSC and Conversation Chronicles), tested cross-domain transfers between human-annotated and synthetic datasets, assessed multiparty dialogue on the Ubuntu IRC benchmark, and conducted human evaluations with eight evaluators.

The evaluation produced four primary findings. First, integrating the proposed framework led to substantial performance improvements across all tested baseline models; for example, on the Conversation Chronicles dataset, fine-tuning an open-source model with the framework increased generation overlap metrics by approximately 50% to 60% compared to standard tuning. Second, the framework outperformed the previous state-of-the-art architecture across multi-session benchmarks. Third, human assessments indicated that topic-based memory retrieval substantially exceeded direct semantic retrieval in both accuracy and recall, translating into marked gains in conversation coherence, fluency, and user engagement. Fourth, cross-domain evaluations revealed strong generalization, as models trained on one dataset retained high performance when evaluated on another with distinct collection properties.

These findings indicate that dialogue systems can achieve long-term coherence without requiring monolithic, bespoke architectures. Organizations deploying conversational systems can adapt existing language models across tasks—such as direct one-on-one chats and complex multiparty conversations—while keeping memory storage and persona management computationally efficient through modular tuning.

Organizations developing customer-facing or companion systems should consider adopting modular memory and persona tracking architectures to improve retention and response quality. Technical teams should implement topic-aware and recency-weighted retrieval rather than relying solely on raw semantic search. Prior to large-scale deployment, practitioners should conduct pilot tests in realistic environments to validate conversational dynamics.

The primary limitation noted in the article is the reliance on synthetic or crowd-sourced multi-session datasets, which may not capture all complexities of authentic real-world dialogues. Additionally, the individual modules currently employ basic summarization and extraction mechanisms, leaving room for further optimization. Consequently, while the framework demonstrates strong technical validity, stakeholders should expect further performance variations when migrating to live production environments.