Paged Context Memory: Runtime Systems for Evidence Blocks, Locality, Speculation, Linking, and Policy Preservation
Description
Paper 1 of this series defined what compiled context is. This paper defines how compiled context lives over time. We argue that AI context systems should be treated as semantic runtimes, not passive stores, and that they require working sets, locality, paging, speculative preparation, lifecycle management, linking, causal provenance, and policy-preserving state transitions.
The paper introduces the evidence block as the fundamental runtime primitive, defines context working sets, semantic locality, and context fragmentation as formal runtime concepts, and extends the runtime with context branch prediction, speculative context execution, context garbage collection, and a context linker. It further formalizes provenance as a causality graph and argues that policy-preserving compilation must preserve policy semantics the way traditional compilers preserve program semantics.
This is Part 2 of the Context Compilation Trilogy, defining the runtime and lifecycle layer for enterprise AI context systems.
Notes
This record is the standalone Zenodo preprint landing page for Part 2 of the Context Compilation Trilogy.
Version note: corrected author affiliation; a prior version listed an incorrect institutional affiliation.
Version 3 (August 2026) — erratum on experimental results. The experimental sections of the version 2 manuscript reported quantitative results that were design targets from the reference architecture, not measured outcomes from completed experiments. Version 3 relabels every such number as a design target or illustrative projection, and distinguishes them typographically from measured results. No measured claims are made in this version. The empirical program that will replace these targets with measured results is public: datasets, runner configurations, and result bundles will be checked into the companion repository as they are produced. The conceptual contributions (representations, taxonomies, algorithms, and formal definitions) are unchanged.