A new arXiv preprint pitches a fix for AI agents that lose the plot across long conversations.
The paper, titled "Disentangling Long-Term Memory via Latent Neuro-Symbolic Reasoning" (arXiv:2609.18461, posted September 17, 2026), introduces a system called LGM. Instead of storing a fixed memory graph or scoring old messages one by one, LGM uses a sparse autoencoder to convert a user's interaction history into latent nodes for each new query, then assigns relational weights between those nodes based on what the query actually needs. A graph encoder runs message passing across that query-specific subgraph, conditioned on the query itself, to produce a memory representation the agent can act on. The paper reports LGM beating unnamed state-of-the-art baselines on long-term personalization benchmarks at capturing both stated preferences and inferred behavioral patterns.
The pitch matters because most "memory" in today's chat agents is either brute-force retrieval, stuffing relevant-looking snippets into a prompt, or a static knowledge graph that does not adapt to what you are currently asking. Neither approach distinguishes an offhand comment from a standing preference, and both get slower and noisier as history piles up. Building the graph fresh, and sparse, for each query is a plausible answer to that scaling problem, if it holds up outside a benchmark.
Take it as an interesting idea, not a shipped product: the preprint lists no named authors or institutional affiliation, has not been peer reviewed, and "significantly outperforms baselines" is the paper's own framing of its own results.