The nearest unsolved problem. Without it every agent starts each day from zero and accumulated experience lives in notes rather than in the system.
About this eventExternal memory and retrieval partly cover the gap but do not accumulate real skill.
Notes, vector retrieval and action history can return a relevant fact to context, but they do not durably change how the task is solved. This is a useful engineering substitute for memory, not learning: without the external record the model remains unchanged and does not automatically transfer the acquired skill.
Source: arXiv · RAGAbout this eventTraining small separate modules — a compromise between frozen weights and full retraining.
Parameter-efficient tuning updates a small part of the system and is therefore cheaper than full retraining. It supports controlled specialisation, but does not by itself solve continual learning: sequential updates still have to be tested for loss of earlier skills.
Source: Microsoft Research · LoRAAbout this eventThe core obstacle: new knowledge erases old, and the model breaks where it worked yesterday.
During sequential training, gradients from a new task can overwrite parameters that matter for earlier tasks. Methods such as EWC protect some of those parameters, but there is still no general solution for large language models learning from an open-ended stream of experience.
Source: PNAS · EWCAbout this eventThere is still no cheap way to prove yesterday's skills survived yesterday's experience.
About this eventA system that after years of work becomes an expert in your specific task rather than the internet average.
Primary material behind this dossier: papers, lab publications and official reports.