[建议 / Feature] Dynamic Context Runtime: Bounded Attention over Unbounded History

提交前确认 · Pre-submission checklist

• 我已搜索过现有 issue，确认这不是重复提议 / I searched existing issues and confirmed this isn't a duplicate.
• 我已阅读 CONTRIBUTING.md / I've read CONTRIBUTING.md.

问题类别 · Category

性能 · Performance (延迟 / 卡顿 / 资源占用)

涉及的 Agent 框架 · Agent framework

多个 / 通用 · Multiple / General

使用场景 · Use case

https://cybersec.org.za/research.html #AI #cyber #security #dev #Africa #LLM #context #rot #cyber
https://cybersec.org.za/papers/dcr-bounded-attention.pdf

Dynamic Context Runtime: Bounded Attention over Unbounded History

Language models degrade as stale and superseded material piles up in their context. Making the window bigger does not fix it. This report describes a runtime that keeps history unbounded and attention bounded — storing everything as immutable spans and a typed provenance graph, then assembling a small working set each turn by solving a knapsack under an explicit token budget. It includes a zero-dependency Rust implementation and an ablation that names which mechanisms are actually carrying the result.
467tokens per query
59×less than full history
7 / 7probes answered
33×history growth, flat cost

建议方案 · Proposal

context rot solution solve context

预期价值 · Expected value

33x context growth flat cost

你认为的优先级 · Your perceived priority

高 · High

你使用的 ZCode 版本 / 环境 · ZCode version / environment

all

补充材料 · Additional context

No response

Start by reading CONTRIBUTING.md and the linked Dynamic Context Runtime paper, including its referenced zero-dependency Rust implementation. The issue names no repository entry point, test, acceptance criteria, or integration target; a concrete scope and verifiable definition of done are needed before implementation can begin.