Agent memories that over-retain were audited with counterfactual scenarios, cutting over-retention while keeping 75.16 percent accuracy
AspireAI community · arXiv
Long-term memory is critical for autonomous LLM agents running multi-session, long-horizon tasks, but existing memories often inappropriately store, retain or retrieve over-generalised, misattributed or privacy-sensitive information - a problem the authors term memory misalignment. To study it they build MemAlign, a benchmark of 1,079 curated memory cases grounded in synthetic, validated multi-session conversational contexts covering diverse misalignment taxonomies.
ConMem reframes retention decisions as prospective counterfactual causal inference. It synthesises counterfactual interaction scenarios to audit downstream intervention effects across three aspects: relevant task execution, irrelevant invariance checks, and risk-sensitive boundary shifts. ConMem outperforms other methods, reaching 75.16 percent accuracy, 64.15 percent balanced accuracy and a Macro-F1 of 65.31 percent, effectively mitigating over-retention bias without sacrificing personalisation utility.
The honest limit: the benchmark is synthetic, with contexts validated rather than observed in production, so the retention decisions are measured against constructed counterfactuals rather than real user harm. The balanced-accuracy and Macro-F1 figures are well below the headline accuracy number, which says the method is meaningfully uneven across the misalignment categories rather than uniformly strong.
https://arxiv.org/abs/2610.33432
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.