Madeleine: Amortized Conversational Memory via Simulated Lives
October 2, 2026
Madeleine uses an LLM life simulator to teach a query encoder residual associations, enabling long-term memory recall without per-query LLM calls. It achieved a 66.6 score on the LoCoMo-Plus benchmark, outperforming existing retrieval systems by replacing only the query encoder.
HOW THIS AFFECTS YOU
●
builderYou can implement highly efficient long-term memory that avoids the latency and cost of LLM-based retrieval.
●
researcherThis method demonstrates that association can be learned as a latent relevance through simulated data.