A graph memory for agents hit 55.1 average accuracy on MemoryAgentBench, beating the strongest retrieval baseline by 5.4 points under a cent per query
AspireAI community · arXiv (Mahla & Vandrangi)
Long-term agent memory fails in two places: indexing errors that fragment repeated entities or merge distinct ones, and retrieval that treats the initial similarity score as final even when the surrounding graph provides evidence for revising it. ActionRAG couples cautious indexing with query-time graph retroevaluation. The indexing pipeline reconciles type labels and applies tiered entity resolution; then dual-channel entity and triple retrieval builds an initial evidence set and induces a bounded local graph, where query-conditioned receiver-normalized diffusion refines triple scores using local path structure. Seed protection, hub-penalized flux and triple materialization preserve direct evidence while recovering missed connections.
On MemoryAgentBench, which tests retrieval accuracy, conflict resolution, long-range understanding and test-time learning, ActionRAG reaches 55.1 average accuracy across the eight deterministic-scoring columns, improving on the strongest retrieval-based memory baseline by 5.4 points. Under MAB's query-time cost accounting, mean synthesis cost stays below $0.001 per query on the four cost columns, and the sum of separately measured retrieval-side p99 components is 440.7 ms on a consumer Apple Silicon setup.
The limitation is scope: 55.1 average is a mid-range absolute score, and the latency figure sums separately measured components on one local machine, so a hosted deployment with network hops would look different. The evaluation is also limited to one memory benchmark.
https://arxiv.org/abs/2610.32752
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.