Personalized agents must reason over long‑term interaction histories to infer both explicit preferences and implicit behavioral cues. Early flat‑retrieval approaches score memory fragments independently, ignoring distributed information, while existing structured memory frameworks rely on query‑agnostic static graphs that cannot capture context‑dependent relations. Raw textual memories are inherently entangled and noisy, making fine‑grained personalization and cross‑session reasoning computationally prohibitive.\
To address this, we introduce LGM, a neuro‑symbolic framework that shifts long‑term memory disentanglement into a continuous latent space. First, we abandon fixed graphs and design a query‑specific latent graph using a sparse autoencoder. This component maps historical interactions into latent memory nodes, disentangles memory traces into sparse concept activations, and dynamically synthesizes query‑aware relational edge weights. Second, a graph encoder treats the query embedding as a conditioning preference, directing non‑linear message passing across a task‑specific latent subgraph, yielding a highly expressive memory representation for activation.\
Extensive experiments on long‑term personalization benchmarks show that LGM significantly outperforms state‑of‑the‑art baselines in capturing both explicit and implicit preferences, enabling more personalized responses.\
Review: LGM’s sparse autoencoder provides query‑driven memory decomposition, and its conditional graph convolution enhances the expressive power of memory retrieval, offering a scalable neuro‑symbolic solution for long‑term personalization.