Howard Chen will present his FPO "Memory Adaptation in Language Models for Long-Horizon and Continual Tasks" on Wednesday, November 19, 2025 at 2:00 PM in CS 105.
The members of Howard’s committee are as follows:
Examiners: Karthik Narasimhan (Co-Adviser), Danqi Chen (Co-Adviser), Tom Griffiths
Readers: Peter Henderson, Jason Weston (Meta)
A copy of his thesis is available upon request. Please email gradinfo@cs.princeton.edu if you would like a copy of the thesis.
Everyone is invited to attend his talk.
Abstract follows below:
Humans can remember events from decades ago and are not overwhelmed by the constant flood of sensory input throughout our lifetime. By contrast, while modern language models (LMs) can memorize massive amounts of data, they fall short in dynamically incorporating the stream of information that emerges over time. As language models operate in a fast-changing world and can now accumulate their own experiences, this limitation becomes acute. It stems from an inflexible memory design: deployed models assume fixed parameters updated only intermittently; larger context windows remain bounded and unevenly effective.
This thesis addresses both forms of inflexibility through memory adaptation: deliberately shaping LM’s memory over time. Two questions guide the work: (1) how to organize in-context (working) memory when accumulated context exceeds what the model can process, and (2) how to update parametric memory with minimal disruption as continual weight changes often induce forgetting.
I first survey how memory is implemented from early to contemporary AI systems and where they fall short. I then proceed in two parts. Part I studies long-horizon tasks in realistic web navigation, showing that standard training does not extend the effective horizon. I introduce a framework that treats long context as a managed data structure the model can traverse. This organization mitigates long-context failures by prioritizing salient information through reasoning and interaction.
Part II studies how to incorporate new data into parameters while avoiding forgetting. I analyze dynamics for two data types—factual knowledge and task specific capability. I then trace failure patterns across stages, and propose mitigation, identifying when forgetting is limited. I also examine tradeoffs between gains in the target domain and drops on other tasks under different learning objectives, and identify conditions with minimal forgetting. Taken together, the results support memory adaptation as a unifying principle: in-context memory organization and deliberate parametric memory updates enable long-horizon and continual tasks.
I close with a roadmap toward self-improving agents: a “machine metabolism” framework where the model decides what to ingest, where to store it, and when to consolidate. I discuss the opportunities and implications for building such a system.