Upcoming TurkuNLP Research Seminar: Learning to Remember 🧠

[Omni Language AI Research @ ELLIS Insititute Finland & TurkuNLP]

LLM agents are increasingly expected to work across interactions that run far past their context window. Good long-term memory takes more than storing extra information. An agent has to learn what to remember, how to organize its experience, and how to assign credit to memory decisions whose effects may only show up much later.

On Friday, 6 November, Sikuan Yan (PhD candidate at LMU Munich and the Munich Center for Machine Learning, advised by Prof. Volker Tresp) will present three works on this problem:

🔹 Memory-R1: memory management and use learned as policies with outcome-driven RL. The agent decides when to add, update, delete or keep information, and how to reason over what it retrieves. 🔹 PyraVid: long videos organized into a coarse-to-fine event hierarchy, for structured retrieval and evidence aggregation at different time scales. 🔹 Memory-R2: RL in environments where the agent's memory keeps changing its future state. It uses local and global group-relative optimization to give fairer credit to individual memory operations.

Taken together, the three works treat memory as an adaptive, structured and trainable part of persistent agents, not a static retrieval pipeline.

📅 Friday, 6 November 2026 🕓 16:00–17:00 (UTC+2) / 15:00 CET 📍 Online via Zoom (Meeting ID: 611 3259 9577) 🔗 Details and calendar invite: https://www.olaresearch.org/seminar/20261106-yan.html

The seminar is open to everyone. Feel free to share it with anyone working on agents, memory or RL.

#LLM #AIAgents #AgentMemory #ReinforcementLearning #MultimodalAI #NLP #TurkuNLP #ELLIS