Much work is currently being done to give AI agents memory abilities without considering the safety risks of doing so. This project will examine the ways episodic memory might make AI agents more dangerous.
About the project
Humans form memories of things we have done and events we have participated in, so-called ‘episodic’ memories. These memories play an essential role in a wide set of mental functions, most importantly our ability to plan our future actions, thereby affecting much of our behavior. When AI agents are deployed and act over long time horizons - months or years - they will need episodic memory abilities to remember what they have done. The safety consequences of agents’ having episodic memory have not yet been widely appreciated or studied.
This project will build off of a research agenda I presented at this year’s Secure and Trustworthy Machine Learning (SaTML) conference in the following paper:
DeChant, 2025. “Episodic memory in AI agents poses risks that should be studied and mitigated” https://arxiv.org/abs/2501.11739
In that paper, I argue that episodic memory will introduce new capabilities into AI agents which will in turn enable new risks. Some of these risks are:
-
Deception: Complex forms of deception would be difficult or impossible without some kind of episodic memory. AI agents acting deceptively will need episodic memory in order to “keep their story straight”: remembering what actually happened, what they claimed happened, to whom they claimed it and when, etc.
-
Improved situational awareness: Endowing an AI agent with episodic memories would allow it to develop a better, more complete picture of the world and its role in it. This would not be limited to knowing whether it was in training, testing, or deployment. It would also develop a knowledge of its own capabilities and limitations that can in some cases only come from observing and later recalling one’s own actions, successes, and failures.
-
The unpredictability of memory: As agents operate over long time horizons, they will lay down memories of events which cannot be predicted in advance. These will then be recalled to inform the agent’s future behavior, at a time and manner which are also difficult or impossible to predict. This unpredictability could make pre-deployment alignment efforts less reliable.
The project will empirically investigate these and possibly other risks. Participants are welcome to suggest others which they think should be studied. We will build a platform and dataset that can be used to understand the risks of episodic memory and that will be helpful in finding ways to mitigate those risks.
For more background on episodic memory in AI, see my paper linked to above and the following paper:
Pink et al, 2025. “Position: Episodic Memory is the Missing Piece for Long-Term LLM Agents” https://arxiv.org/abs/2502.06975
Theory of change
Episodic memory has the potential to make AI agents more dangerous. Many groups are now working on developing episodic memory for AI agents without regard to their safety risks, making this a timely area of study. It is important to understand the risks of episodic memory before such memory is widely deployed in AI agents. Doing so will help shape the direction of episodic memory research and enable the development of safer forms of episodic memory.
Your role
Mentees will help in the creation of a dataset and perform experiments. Beyond that, the degree of autonomy they will have will be determined on a case by case basis. If they would like to, they will be welcome to propose their own ways of implementing the intended research directions or to suggest new directions within the project's scope.
Prerequisites
Highly proficient using python Have conducted previous experiments using PyTorch Experience with at least one LLM API and ideally at least one open source model
Location preference
NA
Application question(s)
-
Please read the first paper mentioned above ( https://arxiv.org/abs/2501.11739 ) and provide a few examples of potentially dangerous behavior that episodic memory could facilitate, either within the categories described there or of a new type that you think of. You are also welcome to suggest some experimental approaches to test these if you'd like.
-
Please also describe a recent experience you have had working with language models. What was the task, what specific methods or code did you use (e.g., API calls, fine-tuning, use of tools)?
About the mentor
Chad DeChant received a Ph.D. in Computer Science in 2025 from Columbia University, where he worked at the intersection of natural language processing and robotics. His dissertation was primarily on remembering, summarizing, and answering questions about robotic actions. He created and taught a course for Columbia undergraduates on “AI Safety, Ethics, and Policy.” He will begin a postdoc position in January at NYU under the direction of Marcelo Mattar, an advisor to the UK AISI whose lab studies memory, planning, and neuro-inspired AI.