Jonas Niederle

Learning agents are increasingly applied to solve complex tasks that require extensive context and, possibly, collaboration between agents. In solving these tasks, agents are still limited by the knowledge acquired during training or provided in a limited working context. Meanwhile, the world changes, as do our requirements for agents and the context in which they operate. How to continuously adapt agents to these changes remains a challenge. Current approaches focus on enlarging the context of single agents or retrieving from growing collections, but lack principled methods for explicitly deciding what to forget, handling conflicting and uncertain information, and sharing memory between agents.

In this PhD, we will develop methods that equip agents with an adaptive memory: the ability to iteratively update, compress, efficiently retrieve and importantly, forget information. A first question is how to efficiently represent memory. Recent work suggests a shared underlying space of concepts might exist, which motivates a memory shared across agents. We will develop mechanisms that explicitly decide what to keep, learning these decisions from data and surrounding context, as recent evidence suggests such learned decisions are effective within single LLMs. Memory must also represent uncertainty about the world, including conflicting information, for which we will explore variational and probabilistic methods.

We will characterize where complexity is required in intelligent systems: in the training of single agents, the tools that we equip them with, their communication or their memory. We hypothesize that an effective memory capability allows smaller, efficient agents to match larger ones that are available only to a few players with large compute resources, enabling more people to benefit from these systems. Adaptive memory should also enable multiple agents to collaborate through a shared memory. We will evaluate the proposed methods on complex tasks in long-horizon settings, where information changes over time. Ultimately, our work paves the way for long-running, efficient agents that operate in a changing world and benefit many people.