New Research Tackles Memory Management and Grounding Challenges in AI Agents

Four new papers address memory organization, attribution, and hardware acceleration for language models and AI agents.

Researchers have published four papers addressing distinct challenges in AI systems, from conversational memory to hardware acceleration.

According to arxiv.org, TrajWiki introduces a trajectory-based memory framework for long-horizon conversational agents. The system represents each memory as a “source-grounded evolution trajectory” maintained through immutable episodic snapshots and claim-level operations including ADD, REVISE, and DEPRECATE. The framework includes a “Memory Wiki” layer designed to reduce fragmentation and retrieval costs.

A separate paper on arxiv.org presents MemArbiter, which addresses what researchers call the “Memory-Action Gap” in LLM agents. The system organizes interaction histories into atomic items across five functional Memory Banks, combining “bank-level demand, item-level relevance, focal-ambient representations, and a temporal presentation gate” to control memory salience. According to the research, MemArbiter achieved success rates of 82.8% and 92.5% under 500- and 750-token budgets in ALFWorld tests, outperforming the strongest baseline by 20.9 and 25.4 percentage points respectively.

On the hardware front, arxiv.org reports that ThAME proposes a 3D heterogeneous multi-chiplet architecture for Mixture of Experts (MoE) inference, achieving speedups of up to 15.7x and energy efficiency improvements of up to 9.8x over state-of-the-art counterparts.

Finally, TRACE-TS, described on arxiv.org, introduces attribution-grounded reasoning for wearable sensor data, achieving 84.43% accuracy and 81.24% F1 scores across seven benchmarks.