ETAMem split agent memory into event, affect and time views, beat the strongest memory baseline on LoCoMo by 9.38 points, and built its memory with 67.92 percent fewer tokens
AspireAI community · arXiv
Long-term agent memory systems tend to mix heterogeneous information inside coarse memory units and then organise everything by semantic similarity, which misses associations that are not about meaning. ETAMem parses conversations into atomic units, separates affective information from the rest, and structures the result into three complementary hierarchical views - Event, Affect and Time - with entropy-based retrieval allocating a query-specific budget across them before searching each view hierarchically.
On LoCoMo the system reaches 84.68 percent overall accuracy, beating TiMem, the strongest baseline it compares against, by 9.38 percentage points and MAGMA by 11.43, with relative gains of 12.46 percent and 15.31 percent on multi-hop questions and 37.51 percent on open-domain ones. On LongMemEval-S it scores 83.00 percent, 7.96 percent above TiMem and 28.00 points above MAGMA, and it leads on multi-session and temporal-reasoning questions by 8.12 and 6.77 points. The efficiency result is as notable: memory construction tokens fall 67.92 percent against TiMem and 88.51 percent against A-MEM, and on LongMemEval-S it needs only 63.41k tokens per session, 88.75 percent fewer than TiMem and 67.70 percent fewer than MAGMA.
Ablations point at the mechanism rather than at scale. Removing affective-memory-unit extraction drops LoCoMo from 84.68 to 82.14 and LongMemEval-S from 83.00 to 80.60, while removing the hierarchical organisation drops them to 82.86 and 80.60. The limits are ordinary for this literature: two long-term conversation benchmarks, baselines chosen by the authors, and no evidence that the same view split transfers to memory-heavy agents operating on tools rather than dialogue.
Read the story and explore more from the AspireAI community.
Two sparse-autoencoder analyses have resisted scaling: matching semantically similar features across layers, and compressing large feature circuits into interpretable supernodes. This paper argues both are the same underlying problem, estimating semantic distances between SAE features that sit on different activation manifolds. Its representation treats each feature not as a single decoder vector but as an activation-weighted distribution over the hidden states expressing it, projects those distributions into a shared reference space, and compares them with Wasserstein distance. The representation is proved invariant to activation rescaling and stable under perturbation, and recovers true matches under finite-sample margin conditions. Empirically it outperformed decoder-vector and LLM-based baselines and captured subtle functional distinctions between related features, and it compressed large feature circuits into interpretable supernodes automatically. The limitation is that this is interpretability tooling: the paper demonstrates better feature matching and circuit compression, not that downstream interpretability claims become more accurate.