TiMem: Temporal-Hierarchical Memory Framework for LLM Conversational Agents
TiMem introduces an innovative memory framework tailored for long-horizon conversational agents, tackling the limitations of finite context windows in large language models (LLMs). It structures dialogues using a Temporal Memory Tree (TMT), facilitating the transition from raw data to refined persona representations. The framework is characterized by three essential features: a temporal-hierarchical structure through TMT, semantic-guided integration for memory consolidation without the need for fine-tuning, and complexity-aware memory retrieval that optimizes both accuracy and efficiency in responses. This strategy seeks to enhance long-term personalization and minimize memory fragmentation.
Key facts
- TiMem is a temporal-hierarchical memory framework for long-horizon conversational agents.
- It uses a Temporal Memory Tree (TMT) to organize conversations.
- The framework enables systematic memory consolidation from raw observations to abstracted persona representations.
- It has three core properties: temporal-hierarchical organization, semantic-guided consolidation, and complexity-aware memory recall.
- Semantic-guided consolidation allows memory integration across hierarchical levels without fine-tuning.
- Complexity-aware memory recall balances precision and efficiency across queries of varying complexity.
- The framework addresses the problem of finite context windows in LLMs.
- TiMem aims to improve long-horizon personalization and reduce fragmented memories.
Entities
Institutions
- arXiv