LLM Memory

The field asks how language models can carry useful information beyond the active context, from facts encoded in model weights to persistent memories for agents. Approaches include retrieval-augmented stores built from summaries, facts, graphs, or key-value records; recurrent or attention-coupled memory modules; and fine-tuning or editing model parameters. Benchmarks increasingly test long conversations, multi-hop and temporal questions, streaming updates, personalization, and forgetting rather than only one-shot recall. A central tension is compressing and updating memory cheaply while preserving relevant detail and the ability to correct stale or misleading information; retrieved memories can also distort reasoning.