TL;DR
ETAMem is a novel multi-view memory framework for Large Language Models (LLMs) that decouples conversational data into three specialized hierarchical views: Event, Affect, and Time. By organizing information based on its inherent characteristics and using an entropy-guided retrieval system, it significantly improves the accuracy and efficiency of long-term memory for AI agents.
What Problem Does This Paper Address?
Building "long-term memory" for AI agents is more than just storing a giant text file of past chats. Current systems struggle with two main issues:
- Information Entanglement: A single sentence like "I went hiking last Wednesday and felt exhausted but happy" contains a factual event (hiking), a temporal context (last Wednesday), and affective states (exhausted, happy). Standard systems often store this as one "blob," making it hard to retrieve just the user's preferences or just the timeline of events without bringing along irrelevant noise.
- Organizational Mismatch: Most memory systems organize data solely by semantic similarity (how similar the words are). However, human-like memory often links things differently—we might remember all our "favorite foods" together (affective) or everything that happened "last summer" (temporal). A one-size-fits-all semantic tree often scatters related preferences across different branches.
Methods
The authors propose ETAMem, which processes dialogue through a three-stage pipeline:
- Heterogeneous Information Parsing: Instead of storing raw dialogue turns, ETAMem uses an LLM to extract Atomic Memory Units (AMUs)—fine-grained, independent facts. It then uses a sentiment classifier to label these units as positive, negative, or neutral.
- Multi-View Memory Structuring: AMUs are routed into three distinct hierarchies:
- Event View: Groups neutral facts using standard textual embeddings.
- Affect View: Groups emotional and preference-based data using specialized affect-aware representations.
- Time View: A fixed hierarchy (Session → Day → Week → Month) that indexes all memories chronologically.
- Dynamic Maintenance & Adaptive Calibration: As the conversation grows, ETAMem automatically splits or merges nodes in these trees. Crucially, it uses "Adaptive Calibration" to adjust the thresholds for these operations based on the actual distribution of the user's data, preventing the memory structure from becoming too bloated or too fragmented.
- Entropy-based Retrieval: When a user asks a question, ETAMem doesn't search all views equally. It calculates a "confidence" score (based on normalized entropy) for each view. If a query is clearly about a feeling (e.g., "What do I like?"), the system allocates more of its retrieval "budget" to the Affect view.
Datasets
The framework was evaluated on three diverse benchmarks:
- LoCoMo: A dataset focused on multi-session dialogues with complex temporal information.
- LongMemEval-S: A benchmark for testing memory over very long-term interactions, including knowledge updates.
- CloneMem: A personal memory benchmark derived from real-world digital traces like diaries and emails.
Key Insights and Ideas
- Decoupling is Key: The "aha moment" is that different types of information require different organizational logic. You can't treat a user's "love for sushi" the same way you treat the fact that they "bought milk on Tuesday."
- Specialized Encoders Matter: The paper shows that using a dedicated affective encoder for the Affect view significantly outperforms using a general text encoder for everything.
- Adaptive Routing Saves Tokens: By using entropy to decide which memory view is most relevant to a query, the system avoids "blind" searching, which saves both time and expensive LLM tokens.
- Fine-Grained "Atomic" Units: Moving from "dialogue turns" to "atomic facts" (AMUs) reduces noise. It’s the difference between searching through a stack of books vs. a well-indexed set of flashcards.
Key Results
ETAMem sets new state-of-the-art (SOTA) records across multiple benchmarks:
- Accuracy: It achieved 84.68% accuracy on LoCoMo (a 12.46% relative improvement over the previous best, TiMem) and 83.00% on LongMemEval-S (a 7.96% improvement).
- Efficiency: Despite its complexity, ETAMem is remarkably efficient. It reduced the number of tokens needed for memory construction by 67.92% compared to TiMem and 88.51% compared to A-MEM.
- Reasoning: The system showed particularly strong gains in multi-hop reasoning (connecting multiple facts) and temporal reasoning, thanks to its dedicated Time and Event views.
Limitations and Future Work
The authors identify a few areas for improvement:
- Sentiment vs. Affect: Currently, the "Affect" view relies on sentiment polarity (positive/negative). This might miss "neutral" preferences (e.g., "I usually take the bus") that are important for personalization but don't carry a strong emotional charge.
- Explicit Relevance: Future versions could benefit from more explicit "affective relevance" classification rather than just relying on sentiment.
- Complex Emotions: Moving beyond simple polarity to more nuanced emotional models (like Joy, Anger, or Sadness) could further refine the Affect view.
Summary updated