AutoViewMem: Self-Configuring Orthogonal Views for... | AI Research

AutoViewMem: Self-Configuring Orthogonal Views for Conversational Long-Term Memory
Large language model (LLM) agents often struggle to maintain long-term memory because they store diverse information—such as personal preferences, event details, and changing constraints—in a single, messy representation. This "semantic interference" makes it difficult for the model to retrieve the right information when needed, as unrelated facts often get mixed together. AutoViewMem addresses this by organizing memory into distinct, focused "views" before it is even stored. By disentangling information at the time it is written rather than during retrieval, the system allows for more accurate and relevant evidence gathering without requiring complex search processes. The ai search story also surfaces in OpenAI Says AI Found Possible Navier–Stokes..., adding another angle.

Organizing Memory Through Semantic Views

Instead of treating every piece of information as a generic entry, AutoViewMem uses a data-driven approach to create "views." A view acts as a specific lens—such as a preference tracker or an event logger—that extracts only the relevant details from a conversation. The system automatically discovers these candidate views from the ongoing interaction, selects a diverse and complementary set using a mathematical method called a Determinantal Point Process (DPP), and uses them to guide the structured extraction of memories. Because these memories are categorized into specific views at the moment they are written, the system can perform a standard, simple search later to find exactly what it needs.

Improving Consistency with Offline Consolidation

While the online extraction process focuses on capturing information quickly, AutoViewMem includes an offline consolidation stage to keep the memory store clean and compact. As the conversation progresses, the system periodically reviews stored items to remove duplicates and merge related facts. It uses a similarity graph to identify overlapping information and employs an LLM to decide whether to combine records or keep them separate based on whether they represent the same event, complementary details, or entirely independent facts. This ensures that the memory remains organized and consistent over long periods. The ai search story also surfaces in Qwen Developers Open-Source Local-First Search Layer..., adding another angle.

Performance and Results

Experiments on benchmarks like LoCoMo and PersonaMem demonstrate that AutoViewMem outperforms traditional memory systems. By using these structured, multi-view representations, the model achieves better results in long-horizon question answering and personalization tasks. The research shows that this approach is particularly effective at helping the model synthesize information and recall specific reasons or preferences, rather than just retrieving raw facts. Furthermore, because the heavy lifting is done during the writing phase, the system maintains a simple and efficient retrieval pipeline, making it a practical solution for long-term agent interactions. The ai search story also surfaces in Stanford Researchers Reportedly Develop Paper2Agent to..., adding another angle. as detailed in the full paper on Arxiv