Engramme – Remember Everything

Overview

Memory is essential to learning and intelligence. Long-term test-time memory is emerging as a bottleneck in current AI, limiting performance in tasks that require models to marshal large amounts of new, often action-dependent and dynamically evolving, information.

The next frontier of AI memory requires scaling to memory that is lifelong, fast, and accurate. Lifelong AI memory will open up qualitatively new abilities, from personalization on decades of user information to contextualizing on weeks to months of agentic trajectories. AI with lifelong memory also promises to alleviate devastating conditions of memory loss and even augment human cognition.

To galvanize work on the frontier of AI memory, this workshop convenes experts from diverse related fields including agentic memory, long-context models, neural memory, personalization, and the cognitive neuroscience of memory. The aim is to foster interdisciplinary dialog toward a unified framework and new datasets and benchmarks appropriate for advancing AI memory.

Why now

Long-term memory is a key open challenge in AI and a central goal in the quest for AGI. Here we define lifelong memory as the scale beyond 100M tokens, informed by a conservative estimate based on decades of a person's documents or days to weeks of agentic trajectories.

There is a disconnect among fields working on long-context models, RAG, tool-based agentic memory, test-time training, LLM finetuning, and the cognitive neuroscience of memory. Further, the community lacks adequate benchmarks for evaluating lifelong memory for AI.

This workshop will facilitate the discussion of unifying principles, comparison of models and benchmarks, and systematic evaluation of the merits and pitfalls of current approaches — bringing forward a much-needed transformation to solve the problem of lifelong memory in AI.