Researchers have published four new papers addressing critical challenges in large language model (LLM) memory management, reasoning, and inference.
According to arxiv.org, TrajWiki introduces a trajectory-based memory framework for long-horizon conversational agents. The system represents memory as “source-grounded evolution trajectories” maintained through immutable episodic snapshots and claim-level operations including ADD, REVISE, and DEPRECATE. TrajWiki includes a “Memory Wiki” layer designed to reduce fragmentation and retrieval costs.
In a separate paper, arxiv.org reports that MemArbiter addresses what researchers call the “Memory-Action Gap” through function-aware memory arbitration. The framework organizes interaction histories into five functional Memory Banks and uses “bank-level demand, item-level relevance, focal-ambient representations, and a temporal presentation gate” to control memory salience. Testing on ALFWorld showed MemArbiter achieved success rates of 82.8% and 92.5% under 500- and 750-token budgets, outperforming the strongest baseline by 20.9 and 25.4 percentage points respectively.
According to arxiv.org, ThAME proposes a three-dimensional heterogeneous multi-chiplet architecture for Mixture of Experts (MoE) inference, using Ferroelectric Field-Effect Transistor-based and DRAM-based memory chiplets. The system demonstrated up to 15.7x speedup and 9.8x energy efficiency improvements over existing counterparts.
Finally, arxiv.org describes TRACE-TS, a framework for “structured and signal-grounded reasoning over wearable time series” that achieved 84.43% average accuracy across seven wearable benchmarks.