A page-based protocol for managing long-horizon LLM context as a logical address space, with demand paging, evidence resolution, and same-origin memory.
一个基于 Page 的长程 LLM 上下文管理协议，将上下文建模为逻辑地址空间，并支持按需分页、证据分辨率和同源记忆。
Paged-Context-Protocol (PCP) 是一种将 LLM 上下文建模为**地址空间（Address Space）**而非单纯线性缓存的上下文治理协议。
PCP 将碎片化的 Token 流与外部数据源转化为离散、可寻址的逻辑页（Logical Pages），并通过递归逻辑树（Logic Trees）、按需下钻和后台整理，在有限上下文窗口内提高召回、溯源和噪声控制能力。
- 逻辑地址空间 (LAS)：将物理存储（PBlock）与逻辑页面（Logical Pages）解耦，支持长历史、文件和仓库的统一寻址。
- 指令驱动寻址：定义
Consult、Explore、Shelve、Purge等指令，用于按需下钻、物理探测、上下文折叠和错误召回隔离。 - 递归逻辑树 (Logic Trees)：通过
Summary、Detail、Unpacked视图在章节、主题或页面层级之间切换。 - 草稿页 (Draft Pages)：在不直接注入全文的情况下，让 Worker 感知候选物理块的逻辑分布。
- 结构化上下文合成：以 XML 或等价结构化格式提供页面边界、ID、时间戳和信任标注，降低上下文混淆风险。
普通长上下文、滑动窗口、一次性 RAG 和滚动摘要都容易在长程任务中丢失低频但关键的逻辑锚点。PCP 的目标是让上下文成为可寻址、可下钻、可折叠、可追溯的系统状态，而不是每轮临时拼接的一段文本。
PCP 目前是协议与工程设计草案。它提出的效果应通过实现和评测验证，例如证据召回率、上下文污染率、摘要损失率、Consult 成功率、成本和延迟。
PCP-native Memory profile 已并入本仓库，见 memory/SPEC.md。它定义了同源持久化 Page Store 的基本契约，包括 QueryMemory、FetchMemory、content_mode、available_modes、版本与溯源。
Paged-Context-Protocol (PCP) is a context governance protocol that models LLM context as an Address Space rather than a linear cache.
PCP transforms fragmented token streams and external data sources into discrete, addressable Logical Pages. It uses recursive logic trees, demand-driven drill-down, and background consolidation to improve recall, traceability, and noise control within a finite context window.
- Logical Address Space (LAS): Decouples physical storage (PBlocks) from addressable logical pages.
- Instruction-Driven Addressing: Defines
Consult,Explore,Shelve, andPurgefor drill-down, physical probing, context folding, and negative feedback. - Recursive Logic Trees: Switches between
Summary,Detail, andUnpackedviews across topic and page levels. - Draft Pages: Exposes the logical distribution of candidate physical blocks without immediately injecting full source content.
- Structured Context Synthesis: Uses XML or an equivalent structured format to preserve page boundaries, IDs, timestamps, and trust labels.
Long-context models, sliding windows, one-shot RAG, and rolling summaries can still lose low-frequency but critical logical anchors in long-horizon tasks. PCP aims to make context addressable, drillable, foldable, and auditable instead of rebuilding it as ad hoc text each turn.
PCP is currently a protocol and engineering design draft. Its claims should be validated empirically with implementation benchmarks such as evidence recall, context pollution, summary loss, Consult success rate, cost, and latency.
The PCP-native Memory profile now lives in this repository. See memory/SPEC.md for the same-origin persistent Page Store contract, including QueryMemory, FetchMemory, content_mode, available_modes, versioning, and provenance.
详细协议规范请查阅 / Please refer to: PROTOCOL.md (CN) | PROTOCOL-en.md (EN) | memory/SPEC.md.