Learning through experience:Episodic memory representation for cognitive agents
Keywords: Episodic Memory, Bio inspired Robot learning, incremental Memory structures
Abstract: As the demand for intelligent robots and cognitive agents rises, the ability to retain and utilize past experiences through episodic memory has become crucial, especially for social companion robots that rely on previous interactions for task execution. To address this, we introduce Episodic Memory for Cognitive Agents (EMCA), a novel framework that advances knowledge representation by integrating real-world interactions. EMCA enables agents to adapt to complex environments by learning from tasks, interacting with humans, and processing multimodal data—such as speech, vision, and non-verbal cues—without pre-training on specific scenarios.
EMCA models episodic memory through a graph-based structure , allowing for incremental storage and retrieval of experiences. Each interaction or event enriches the memory graph, supporting continuous learning and adaptation without extensive retraining. This human-like memory formation optimizes the agent’s ability to retrieve relevant information for tasks like localization, planning, and reasoning based on prior experiences.Unlike conventional models relying on temporal markers or recurrent patterns, EMCA encodes data like human memory, allowing reasoning across diverse scenarios regardless of temporal patterns. The framework dynamically builds a memory graph with semantic and temporal connections based on the agent’s experiences, promoting flexible temporal reasoning. It also introduces mechanisms for clustering new memories and a dynamic retrieval policy that adjusts based on context or query type, ensuring robustness even in unpredictable scenarios. Empirical tests show EMCA adapts effectively to real-world data, offering reliability and flexibility in dynamic environments.
Primary Area: transfer learning, meta learning, and lifelong learning
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.
Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.
No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.
Submission Number: 11037
Loading