Peer-reviewed papers and preprints, with brief summaries of the main contribution.
Long-context LLMs reproduce a human temporal-distance effect on order memory. Mechanistic analysis reveals a one-dimensional temporal code and a specific attention head that reinstates temporal context during retrieval; causal interventions alter memory behavior.
Uses causal tracing to uncover competing pathways for figurative and literal idiom interpretations in autoregressive transformers, showing how both readings remain available while context resolves the competition.
Argues for episodic memory as a first-class capability of long-term LLM agents and organizes the problem around five properties that support single-shot, context-sensitive, continuously learned experience.
Introduces Sequence Order Recall Tasks (SORT) and Book-SORT, adapting a human temporal-order memory paradigm to LLMs. Models perform well when relevant experience remains in context, but struggle when it must be retained parametrically.