A large-scale evaluation found many dedicated agent memory systems underperform naive RAG, and the authors named the gap memory tax - the evidence that memory construction destroys or obscures along with what it reorganizes
AspireAI community · arXiv (Shanghai AI Lab / SJTU)
Memory Tax, an arXiv paper submitted October 25, 2026, asks a question the agent-memory field has largely skipped: what is the net value of building memory versus retrieving directly from raw conversation history? Across multiple benchmarks and model backbones, the authors find that many dedicated memory systems underperform naive RAG in many cases, indicating a negative net value, with some underperforming the naive baseline despite substantially higher costs. They decompose that value into memory benefit and memory tax, attributing both to how memory changes the organization and accessibility of raw information: changes designed to help, such as reorganizing facts or distilling preferences, can also remove or obscure evidence. The paper focuses its analysis on the tax side, studying how it arises from the system and evaluation side and generalizing the concept to other application scenarios, and the authors say they will open-source the code and results. The honest limitation is that the evaluation covers a subset of benchmarks and model backbones by the authors' own admission - they evaluated 20% of PersonaMem-v2, 10% of MINT-Eval and 20% of LongMemEval-s questions for GPT-5.6, and a 10% stratified sample of 60 of 595 corpora - so the negative-net-value finding is directional rather than exhaustive.
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.