From Memory to Autonomous Research: Building Sustainable Long‑Horizon AI Agents

The MLNLP academic talk on July 18, 2026 features Hu Yuyang, a first‑year PhD student at Renmin University’s Gaoling AI Institute, who will present “From Memory to Autonomous Research: Building Sustainable Long‑Horizon AI Agents.”

Hu’s research focuses on long‑horizon agents, specifically on three pillars: context management, agent memory, and autonomous research. He has published multiple papers at top venues such as ACL and ICML, and his open‑source projects (Arbor, CompassMem, SAM) have earned over 3 k GitHub stars.

Abstract : The talk traces the evolution from single‑turn QA assistants to agents capable of autonomous scientific research, emphasizing that long‑horizon tasks require not only reasoning but also effective context handling, persistent memory, and self‑directed exploration. Three representative works are discussed:

SAM introduces a state‑adaptive memory mechanism that dynamically manages and folds context based on task state, mitigating context explosion and information loss during long‑range inference.

AgentFugue expands agent breadth through shared group context, enabling multiple agents to collaborate via a common contextual pool.

CompassMem organizes an agent’s history as an event‑centric, searchable memory graph, allowing the agent to retrieve and reason over past experiences during extended searches.

Arbor targets general autonomous research by employing hypothesis‑tree refinement, enabling the agent to propose, validate, and iteratively improve research hypotheses without human intervention.

Collectively, these contributions demonstrate how systematic improvements in context, memory, and autonomous reasoning can empower AI agents to operate reliably over long time horizons in real‑world environments.