The Context Coordinate System and the Principle of Unbroken Memory (컨텍스트 좌표계와 끊기지 않는 기억의 원리)

The fundamental physical constraint of long-horizon large language model (LLM) agent systems is that the Active Context Window is finite. Prevailing agent frameworks rely on Summary-based Compaction, overwriting the existing dialogue history with a generative summary. However, this destroys the return address to the original log, provides no conservation check for omitted items, and severs the Continuity of the Present in the middle of ongoing work. This paper proposes Context Lattice, an address-based memory management architecture built on the Context Coordinate System, Reference-based Folding, and Asynchronous Stepped Window Shifts:(1) Immutable Raw Log & 5-Level Canonical Coordinates with mathematical Conservation Equations enforced by deterministic runtime code;(2) 5-Slot Frequency-Ordered Window Partitioning and Tier-1 Folding of completed tool payloads (85–95% of context volume) into 1-line coordinate pointers with Append-at-Tail lossless dereferencing;(3) Asynchronous Pre-building and Idle-Boundary Atomic Stepped Shifts to preserve prompt prefix (KV) cache stability and unbroken working memory. Across four quantitative studies using real agent development transcripts and our open-source deterministic coordinate engine (context-lattice), canonical coordinate dereferencing achieves 100.0% recall on 20 ground-truth probes (vs. 30.0% on compaction summary alone and 40.0% with kept pre-compaction messages), reduces context volume by 79–91% with 0% dialogue loss (~8.6–9x longer single-window span), achieves a 95.9% prompt prefix cache hit rate over 100 turns, and processes 50,000 log lines (10,000 turns) in ~185–260 ms with 0% integrity error. --- [한국어 초록]본 논문은 기존 덮어쓰기 요약 압축(Summary-based Compaction)이 야기하는 역참조 주소 상실, 무결성 검증 부재, 작업 중 '현재 맥락의 단절' 문제를 해결하기 위해, 5계층 정준 컨텍스트 좌표계(Context Coordinate System)와 참조 기반 도구 폴딩(Reference Folding), 휴지기 계단식 윈도우 전환(Stepped Window Shift)에 기반한 기억 관리 구조(Context Lattice)를 제안하고 실증한다. (한국어 및 영어 전문 PDF, 1분 개념 영상 수록)