Contributed Talk Sessions | Poster Sessions | All Posters | Search Papers

Poster Session C: Friday, August 15, 2:00 – 5:00 pm, de Brug & E‑Hall

A Unified Benchmark for Human-Like Memory in Artificial Agents

Lucas Gruaz¹, Aude Maier, Johanni Brea²; ¹EPFL - EPF Lausanne, ²Swiss Federal Institute of Technology Lausanne

Presenter: Lucas Gruaz

Human memory exhibits a diverse range of well-documented phenomena, including forgetting curves, interference effects, and schema-based distortions. While existing computational models attempt to capture aspects of these phenomena, they are often evaluated in isolation using task-specific experimental setups, limiting their generalizability and comparability. We develop a unified benchmark for systematically evaluating memory models based on their ability to reproduce human-like memory phenomena. Our approach includes: (1) analyzing and formalizing a diverse set of memory phenomena in generalizable terms, independent of specific experimental paradigms, and (2) developing an evaluation framework that tests these phenomena within a common environment. This allows to test all phenomena on the same memory-augmented agent. We test different memory models on schema-based distortion, memory conjunction errors, contiguity and recency effects. We find that none of the tested memory models qualitatively matches human memory behavior on all these four phenomena, and we identify promising directions for future research on memory models.

Topic Area: Memory, Spatial Cognition & Skill Learning

Extended Abstract: Full Text PDF