# Blast Radius

Source: [arXiv](https://arxiv.org/abs/2608.07440v1)  
Feed7 permalink: https://feed7.dev/p/2608-07440v1-1gi5t78  
Published: 2026-08-07T17:23:53.000Z  
Trust: Needs Review (needs_review)

## Why Included

Blast Radius predicts which code and context a prompt will reach, then reversibly archives unused history. Across seven OpenAI models, it cut token consumption by 17–26%.

## Source Summary

Blast Radius manages coding-agent memory by estimating a prompt’s reach across context and code. Across **seven OpenAI models**, it reduced token use by **17–26%** while keeping archived context byte-exact and reversible.

## Practical Implication

For long-running agents, consider eviction policies that archive rather than summarize context, and track repeated transcripts separately. The paper classified **378 of 450** evicted records as recurring dead matter.

## Agent-Ready Context

Blast Radius manages coding-agent memory by estimating a prompt’s reach across context and code. Across **seven OpenAI models**, it reduced token use by **17–26%** while keeping archived context byte-exact and reversible.

For long-running agents, consider eviction policies that archive rather than summarize context, and track repeated transcripts separately. The paper classified **378 of 450** evicted records as recurring dead matter.

The evidence comes from the authors’ tested policies and models. **Zero archived records were recalled**, but that observation does not establish how eviction behaves across other repositories, workflows, or longer horizons.

## Connected Context

Feed7 judgment across 409 accumulated Signals:

This narrows long-context memory management to a reversible eviction problem: repeated prompt debris can be removed without lossy summarization or deleting its source record. It complements durable shared-memory proposals but does not show that archived material is useful later—the absence of recalls leaves retrieval policy, provenance, and long-horizon behavior unresolved.

- [Chained Recursive Language Models for Multi-Iteration Reasoning](https://feed7.dev/p/2608-05124v1-05haobv) — Chained RLM preserves summaries between fresh contexts, whereas Blast Radius archives evicted records byte-exactly; together they expose a tradeoff between compact carried state and reversible source preservation.
- [TencentCloud/TencentDB-Agent-Memory](https://feed7.dev/p/tencentdb-agent-memory-0gx8nnn) — The memory hub provides a governed destination for reusable context, while Blast Radius addresses which prompt records should leave active context; combining the ideas would require an explicit recall policy that Blast Radius did not exercise.
- [WTF Is the Context Layer? The Missing Infrastructure for Production Agents — Prukalpa Sankar](https://feed7.dev/p/wtf-is-the-context-layer-the-missing-infrastructure-for-production-agent-0t47xqf) — Shared, versioned context separates durable organizational knowledge from an individual agent prompt, reinforcing Blast Radius’s distinction between preserving a record and keeping it continuously in active context.

## Context Map

- Layer: context
- Domains: coding
- Topics: context-engineering, context-caching, agent-memory

## Uncertainty

- The evidence comes from the authors’ tested policies and models. **Zero archived records were recalled**, but that observation does not establish how eviction behaves across other repositories, workflows, or longer horizons.

## 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.