# Eviction as Estimation: A Fixed-Lag Smoothing View of Test-Time Memory, and When Measuring Beats Accumulating

Source: [arXiv](https://arxiv.org/abs/2607.24667v1)  
Feed7 permalink: https://feed7.dev/p/2607-24667v1-177qf29  
Published: 2026-07-27T17:08:27.000Z  
Trust: Needs Review (needs_review)

## Why Included

A delayed KV-eviction policy helps when memory reuse is sharp and observable, but loses its edge on natural-text benchmarks—a useful warning for agent-memory claims.

## Source Summary

RMM delays eviction long enough to observe which stored items support a correct near-future prediction. It is a training-free generalization of **H2O** and reduced to H2O when that utility measurement was uniform.

## Practical Implication

For builders tuning bounded agent memory, test eviction policies on the actual reuse pattern. RMM made small memory act much larger in controlled delayed-reuse settings, but was only **on par with H2O** for single-turn QA and trailed H2O and SnapKV in streaming multi-turn tests.

## Agent-Ready Context

RMM delays eviction long enough to observe which stored items support a correct near-future prediction. It is a training-free generalization of **H2O** and reduced to H2O when that utility measurement was uniform.

For builders tuning bounded agent memory, test eviction policies on the actual reuse pattern. RMM made small memory act much larger in controlled delayed-reuse settings, but was only **on par with H2O** for single-turn QA and trailed H2O and SnapKV in streaming multi-turn tests.

On natural text, correctness weighting changed attention too little, so demonstrated utility largely collapsed into accumulated attention. The proposed framework may help classify workloads, but the reported policy is not a new general-purpose state of the art.

## Context Map

- Layer: benchmark
- Domains: None
- Topics: agent-memory, context-caching, benchmark-integrity

## Uncertainty

- On natural text, correctness weighting changed attention too little, so demonstrated utility largely collapsed into accumulated attention. The proposed framework may help classify workloads, but the reported policy is not a new general-purpose state of the art.

## Agent Instruction

Use this item as source-backed context. Do not invent claims beyond the linked source. If this item conflicts with another source, call out the conflict.