Madeleine: Learning Involuntary Recall for Conversational Memory from Simulated Lives
A long-term assistant must recall memories whose cues are associatively related to, yet not semantically similar to, the current user turn. Madeleine reframes this involuntary recall as a learnable relevance—the pointwise mutual information of memories under simulated life trajectories—and distills it into a query-side residual adapter, $s_\theta(q,m)=\cos(E(q),E(m))+\cos(E_\theta(q),E(m))$. The system is trained offline on synthetic (trigger, cue) pairs written by an LLM life simulator, and at runtime it plugs into existing vector memories with zero LLM calls, cutting answer context by roughly $20\times$ while setting a new state of the art on LoCoMo-Plus.
The paper offers an elegant System-1 alternative to costly System-2 memory retrieval and delivers strong LoCoMo-Plus numbers, but its empirical scope is narrow and several headline plug-in gains partly reflect backbone upgrades rather than the learned association itself.
The efficiency story is rigorous: standalone Madeleine reaches HyperMem's released score with zero LLM calls and about 1/21 of its answer-input tokens, which matters when retrieval runs at every turn. The query-side-only residual design is well justified by Proposition 1 and the ablations—freezing the memory index lets existing systems adopt Madeleine without re-encoding their banks, and training the memory side or dropping the residual sum both hurt associative recall.
A sizable fraction of the plug-in gains comes from fixing backbone mismatches rather than from learned association: inside T-Mem, switching from bge-m3 to the untrained Qwen3-Embedding-4B raises the score from 33.7 to 53.9, leaving only 6.0 points for Madeleine; inside HyperMem, adding the omitted query instruction accounts for 7.7 of the 13.7-point gain. The end-to-end evaluation is also limited almost entirely to LoCoMo-Plus (401 items), and the harmlessness claim weakens on the 8B backbone, which drops ordinary LoCoMo QA by 2.6 points—a result reported only in Appendix D.
The paper fairly reruns all baselines under the official LoCoMo-Plus protocol with the same reader and judge, so the numbers are internally comparable. The evidence supports the central claim that a residual query adapter outperforms read-time LLM rewriting and generic reasoning retrievers on associative recall: every alternative route from HyDE and QueryLink to reasoning-oriented retrievers and rerankers falls significantly below Madeleine. The authors are transparent that prior published numbers from T-Mem and JustMem were obtained under non-comparable protocols and therefore excluded from the main comparison.
Hyperparameters, simulator prompts, and decontamination rules are documented with unusual care in the appendices, down to LoRA rank, temperature, and the exact negative masking threshold. The main obstacle to independent reproduction is the absence of an explicit code or training-data release in the provided text; reproducing the three-source life-simulator pipeline and the exact same-source negative sampling from the description is feasible but would be labor-intensive without released artifacts.
A long-term conversational assistant must recall the right memory at the right moment, yet the memory that matters most is often not similar to what the user says now. Current systems recover such associations by letting an LLM reason at write or read time, at a cost of hundreds to over a thousand LLM calls per memory bank and up to several thousand context tokens per query. We argue that association is a learnable relevance: the pointwise mutual information of memories under how human lives unfold. We introduce Madeleine, which learns amortized association: offline, an LLM life simulator writes simulated lives, whose cue-trigger pairs teach a query encoder a residual association on top of frozen similarity; online, it calls no LLM and plugs into any vector memory by replacing only the query encoder. On LoCoMo-Plus under the official protocol, Madeleine (I) reaches 66.6 when plugged into HyperMem, the highest among all systems evaluated under this protocol; (II) used alone, reaches the score of HyperMem as released (52.4 vs. 52.9) with zero LLM calls and about 1/21 of its answer context; and (III) lifts T-Mem by 26.2 points, significantly outperforms the same untrained backbone inside both systems, and leaves ordinary QA intact on the 4B backbone.
Pick a starting point or write your own. Challenges run in the background, so you can keep reading while the AI investigates.
No challenges yet. Disagree with the review? Ask the AI to revisit a specific claim.