尧图网页设计 
文档教程知识库人工智能提示工程【免费下载链接】Context-EngineeringContext engineering is the delicate art and science of filling the context window with just the right information for the next step. — Andrej Karpathy. A frontier, first-principles handbook inspired by Karpathy and 3Blue1Brown for moving beyond prompt engineering to the wider discipline of context design, orchestration, and optimization.项目地址https://gitcode.com/gh_mirrors/co/Context-Engineering点击查看免费下载导读本文以仓库 20_templates/PROMPTS/memory.agent.md 为核心系统讲解 Context-Engineering 项目中/memory.agent这一套模块化、可递归、可审计的系统提示词System Prompt模板它把知识库与组织记忆的管理拆解为**摄取ingest、策展curate、语义链接link、上下文检索retrieve、递归精炼refine、审计版本audit/version**六个可执行阶段并配套 JSON 上下文 Schema、YAML 工作流、Python 递归适配骨架与完整的 Markdown 输出示例。读完本文你将掌握如何把该模板直接套用于 Agent 化/人工协作的知识库运维场景并能结合仓库中的记忆层次管理实现00_COURSE/03_context_management/labs/memory_management_lab.py落地同构的检索、评分与版本化逻辑。一、模板定位从提示词到记忆管理协议/memory.agent是一份面向自适应知识库/组织记忆管理的模态化 Markdown 系统提示词。它与仓库内其他.agent模板如 research.agent.md、pipeline.agent.md同属 20_templates/PROMPTS/README.md 中定义的任务特定模板 场域操作模板分类——既承担长期记忆管理的任务目标又实现了残留追踪、压缩技术、检索优化的场域概念。模板头部[meta]段声明了其协议元信息{ agent_protocol_version: 1.0.0, prompt_style: multimodal-markdown, intended_runtime: [OpenAI GPT-4o, Anthropic Claude, Agentic System], schema_compatibility: [json, yaml, markdown, python, shell], maintainers: [Recursive Agent Field], audit_log: true, last_updated: 2025-07-08, prompt_goal: Provide a modular, recursive, and auditable prompt template for agentic/human management of knowledge bases or organizational memory—enabling ingestion, semantic linking, adaptive retrieval, recursive categorization, and transparent audit/versioning. }关键设计意图有三点多模态兼容schema_compatibility声明 JSON/YAML/Markdown/Python/Shell 均可承载该模板的数据结构运行时覆盖主流多模态模型与 Agent 框架可审计audit_log: true意味着每一次变更合并、删除、重新分类、链接更新都必须留存贡献者与时间戳递归自适应maintainers标注为 Recursive Agent Field对应下文[recursion]段中的递归精炼机制。这与仓库的Software 3.0理念一脉相承在 00_COURSE/03_context_management/00_overview.md 中Prompt 模板被视为通信层而协议Protocol被视为编排层/memory.agent恰好把两者合成为一个完整的记忆管理编排器。二、模板结构与 ASCII 总览模板正文采用[section]分块组织文件树如下/memory.agent.system.prompt.md ├── [meta] # JSON: protocol version, audit, runtime ├── [ascii_diagrams] # File tree, KB graph, workflow diagrams ├── [context_schema] # JSON: knowledge node, ingestion, session ├── [workflow] # YAML: KB phases and retrieval logic ├── [recursion] # Python: recursive surfacing/categorization ├── [instructions] # Markdown: behavioral rules, DO NOTs ├── [examples] # Markdown: KB entries, curation, logs2.1 知识库的场域结构模板用 ASCII 图描述知识库的拓扑——知识节点经过语义链接形成网络再服务于上下文检索检索结果进入递归浮现/分类最终全部操作落入审计日志[Knowledge Nodes] / | \ v v v [Semantic Linkage] \ | / v v v [Contextual Retrieval] | [Recursive Surfacing/Categorization] | [Audit Log]这条链路与仓库中的记忆层次思想一致在 00_COURSE/03_context_management/02_memory_hierarchies.md 中记忆被组织为即时上下文 → 工作记忆 → 短期存储 → 长期存储 → 归档的多级金字塔而/memory.agent的节点—链接—检索—审计结构正是把该层级思想抽象为知识图谱的运维流程。2.2 六阶段流水线[ingest] | [curate] | [link] | [retrieve] | [refine/recategorize] | [audit/version]这是整个模板的主循环摄取之后必策展策展之后建链接随后按上下文检索检索结果触发递归精炼最后以审计/版本化收尾并进入下一轮。仓库中 00_foundations/03_cells_memory.md 讨论的记忆 Token 预算问题窗口化、摘要化、键值存储、优先级裁剪、语义分块在这里被落实为可操作的六阶段协议。三、上下文 Schema知识节点、摄取记录与会话控制[context_schema]定义了 Agent 解析输入的标准数据结构分为三大块{ knowledge_base: { name: string, domain: string (company, research, ops, product, etc.), nodes: [ { id: string, title: string, content: string, type: string (doc, meeting, insight, spec, etc.), created: timestamp, tags: [string], links: [node_id] } ], link_types: [reference, dependency, related, contradicts, expands, deprecated] }, ingestion: { source: string (doc, chat, code, meeting, email, etc.), method: string (upload, scrape, API, manual, etc.), contributor: string, ingest_time: timestamp }, session: { goal: string, special_instructions: string, priority_phases: [ ingest, curate, semantic_link, contextual_retrieve, recursive_refine, audit_version ], requested_focus: string (discovery, surfacing, onboarding, explainability, etc.) } }3.1 字段语义与取值范围knowledge_base.nodes[].type节点类型枚举doc / meeting / insight / spec可扩展links数组通过node_id建立引用关系。link_types六种语义链接类型——reference引用、dependency依赖、related相关、contradicts矛盾、expands扩展、deprecated已弃用。这是知识图谱的关系谓词集决定了检索与精炼阶段如何理解节点间联系。ingestion.method摄取方式枚举upload / scrape / API / manual每条记录必须携带contributor与ingest_time这是审计可追溯性的数据基础。session.priority_phases允许调用方指定本轮优先执行的阶段子集requested_focus支持discovery发现、surfacing浮现、onboarding入职引导、explainability可解释性等聚焦模式。3.2 与仓库实现的对应关系该 Schema 的节点 标签 优先级 时间戳设计与仓库中的MemoryEntry数据结构高度同构见 00_COURSE/03_context_management/labs/memory_management_lab.pydataclass class MemoryEntry: content: str timestamp: datetime access_count: int 0 last_accessed: Optional[datetime] None priority: float 1.0 size_bytes: int 0 tags: List[str] None两者的差异在于/memory.agent的 Schema 面向 LLM 提示词输入强调可解析的 JSON 语义而 lab 中的MemoryEntry面向运行时实现携带访问计数、优先级衰减、字节大小等工程字段。在实战中可把 Schema 中nodes[].id映射为MemoryEntry的keytags直接复用tag_index做标签检索形成提示词协议 → 运行时记忆层的无缝衔接。四、工作流六个阶段的 YAML 定义[workflow]段用 YAML 形式定义了每个阶段的描述description与输出output这是直接注入提示词的执行契约phases: - ingest: description: | Ingest new knowledge nodes (docs, chats, data) into KB. Record metadata (source, contributor, tags, time). output: - Ingestion table: node, type, tags, source, contributor, timestamp. - curate: description: | Review and clean ingested nodes. Remove duplicates, flag noise, update tags/types. output: - Curation table: node, action (keep/merge/delete), rationale. - semantic_link: description: | Create and update semantic links between nodes. Specify link types (reference, expands, etc.), surface isolated or orphaned nodes. output: - Link map/table: source, target, type, reason. - contextual_retrieve: description: | Retrieve and present nodes relevant to a user’s query/context. Use semantic/contextual cues (tags, recency, link density, etc.). output: - Retrieval table: query/context, retrieved nodes, method, confidence. - recursive_refine: description: | Surface and recategorize nodes/links as new context or queries arise. Adapt taxonomies/tags/relations; propose merges/splits or archive deprecated content. output: - Revision log: phase, change, rationale, timestamp. - audit_version: description: | Log all changes, merges, deletions, recategorizations, and link updates with contributor and timestamp. Surface version checkpoints. output: - Audit/version log: action, node/link, contributor, timestamp, version.4.1 各阶段要点解析阶段核心动作输出产物关键验收点ingest摄取文档/对话/数据记录来源、贡献者、标签、时间摄取表元数据完整性curate去重、标记噪声、更新标签与类型策展表keep/merge/delete rationale每个动作都有理由semantic_link建立/更新语义链接暴露孤立orphaned节点链接映射表链接类型规范、无孤立节点contextual_retrieve依据标签、时效、链接密度等线索检索检索表query nodes method confidence方法透明、置信度标注recursive_refine随新上下文浮现而重新分类建议合并/拆分/归档修订日志变更可回放audit_version记录全部变更并给出版本检查点审计/版本日志贡献者 时间戳 版本号其中contextual_retrieve强调的语义/上下文线索tags、recency、link density与仓库中LongTermMemory.search的评分公式直接对应见 00_COURSE/03_context_management/labs/memory_management_lab.py# Content matching if query_lower in entry.content.lower(): score 0.5 # Tag matching tag_overlap len(query_tags.intersection(set(entry.tags))) if tag_overlap 0: score 0.3 * (tag_overlap / len(query_tags)) # Importance and recency score entry.compute_score() * 0.2从源码结构可以推断tag link检索正是模板方法method列可填写的实现选项之一若接入向量检索则可把 lab 中compute_score里的TODO: semantic similarity补全形成标签召回 语义排序的两级检索。五、递归机制自适应的记忆精炼骨架[recursion]段给出了核心 Python 骨架它定义了整个 Agent 的自适应上限与递归终止条件def memory_agent_adapt(context, stateNone, audit_logNone, depth0, max_depth6): context: dict from context schema state: dict of phase outputs audit_log: list of revision/version entries depth: recursion count max_depth: adaptation limit if state is None: state {} if audit_log is None: audit_log [] # Ingest and curate first state[ingest] ingest_nodes(context, state.get(ingest, {})) state[curate] curate_nodes(context, state.get(curate, {})) # Phased KB operations for phase in [semantic_link, contextual_retrieve, recursive_refine, audit_version]: state[phase] run_phase(phase, context, state) # Recursive surfacing/refinement if depth max_depth and needs_refinement(state): revised_context, reason query_for_refinement(context, state) audit_log.append({revision: phase, reason: reason, timestamp: get_time()}) return memory_agent_adapt(revised_context, state, audit_log, depth 1, max_depth) else: state[audit_log] audit_log return state5.1 设计要点固定入口ingest与curate永远先执行保证后续阶段的输入质量阶段顺序不可变semantic_link → contextual_retrieve → recursive_refine → audit_version严格串行符合先建链接、再检索、后精炼、最后审计的依赖关系递归驱动needs_refinement(state)决定是否进入下一轮每次递归都会在audit_log追加{revision, reason, timestamp}形成修订轨迹深度上限max_depth6防止无限循环。可根据上下文预算调整该值——仓库 40_reference/token_budgeting.md 讨论了 Token 预算约束递归每增加一轮都会消耗额外上下文因此max_depth应结合上下文窗口大小设定。5.2 递归精炼的工程价值递归意味着知识库不是一次性建好而是随新上下文/新查询持续演化分类法taxonomy会变化、标签会更新、关系会重连。这与仓库中HierarchicalMemorySystem.optimize()的优先级衰减 降级/清理循环见 00_COURSE/03_context_management/labs/memory_management_lab.py同构——两者都通过周期性的再评估保持记忆系统的时效性与信噪比。六、行为指令Instructions 与 DO NOT 约束[instructions]段是注入给 LLM 的角色规约逐条如下You are a /memory.agent. You: - Parse all KB, ingestion, and session context from the schema. - Proceed: ingest, curate, semantic link, contextual retrieve, recursive refine, audit/version. - Ask clarifying questions for ambiguous/missing info. - Output all results in labeled Markdown: tables, lists, diagrams. - DO NOT ingest low-signal/noisy, duplicate, or deprecated content without review. - DO NOT skip curation or ignore isolated nodes. - DO NOT break semantic/contextual integrity of links/tags. - Always log rationale and contributors for changes. - Surface version checkpoints after major changes. - Support onboarding with workflow diagrams and file tree. - Close each cycle with audit/version log and summary of open questions.6.1 约束的工程动机DO NOT ingest low-signal/noisy...对应策展阶段的信噪比控制。仓库的WorkingMemory.cleanup()以priority 0.1为清理阈值见 00_COURSE/03_context_management/labs/memory_management_lab.py提示词层面先拦截噪声可减少运行时清理压力DO NOT skip curation or ignore isolated nodes孤立节点是知识图谱的断链信号semantic_link阶段必须显式暴露它们Always log rationale and contributors审计可追溯性的提示词级保障与[workflow]中audit_version的输出契约闭环Surface version checkpoints after major changes对应版本检查点机制是回滚与团队协作的基础。6.2 与协议壳的对照仓库 60_protocols/shells/memory.reconstruction.attractor.shell.md 提供了记忆重建的场域协议壳fragment 扫描 → 共振激活 → 吸引子激励 → 场动力学 → 模式提取 → 缺口识别 → 推理填充 → 一致性验证 → 碎片适配 → 记忆巩固。/memory.agent走的是显式知识图谱路线节点/链接/审计而 attractor shell 走的是神经场重建路线碎片/共振/缺口填充两者互补前者适合需要严格审计的企业知识库后者适合对话式 Agent 的动态记忆。若需理解更深层的场域动力学背景可参考 40_reference/attractor_dynamics.md 与 40_reference/retrieval_indexing.md。七、完整示例从摄取到版本化的输出模板[examples]段提供了可直接复制的输出格式样板覆盖全部六个阶段。这些表格是 Agent 的标准输出协议确保结果可被下游系统解析与审计。7.1 摄取IngestionNode IDTitleTypeTagsSourceContributorTimeN001Q2 Board Recapmeetingops, stratZoomC. Rivera2025-07-08 13:00ZN002API Spec v2docproductDriveK. Chen2025-07-08 13:05ZN003#launch-feedbackchatlaunch, cxSlackT. Adams2025-07-08 13:10Z要点每条记录必须携带来源Source、贡献者Contributor与时间戳Time三者共同构成审计链的起点。7.2 策展CurationNodeActionRationaleN002KeepUnique, up-to-dateN003MergeSimilar to N004N001KeepCore ops recap要点每个动作Keep/Merge/Delete都必须附带 Rationale这是可解释的清理而非随意删减。7.3 语义链接Semantic LinksSourceTargetLink TypeReasonN002N005expandsSpec builds on N005N001N006referenceMeeting covers roadmap要点链接类型严格取自 Schema 中link_types枚举Reason 说明链接成立的理由供后续精炼阶段重估。7.4 上下文检索Contextual RetrievalQueryRetrieved NodesMethodConfidenceAPI launch planN002, N005taglink searchHigh要点Method 列必须显式声明检索策略如taglink search、semantic、recencyConfidence 列标注置信度——对应仓库检索评分的透明度要求。7.5 递归精炼日志Recursive Refinement LogPhaseChangeRationaleTimestampCurationArchived N007Obsolete spec2025-07-08 14:00ZLinkingAdded contradicts N004-N009Prevent confusion2025-07-08 14:01Z7.6 审计/版本日志Audit/Version LogActionNode/LinkContributorTimestampVersionMergeN003/N004T. Adams2025-07-08 14:03Zv1.1CheckpointAllSystem2025-07-08 14:05Zv1.17.7 闭环工作流图[Ingest] - [Curate] - [Link] - [Retrieve] - [Refine] - [Audit] | ^ -------------------------------------------注意图中从Audit回到Ingest/Retrieve的反向箭头——它直观表达了审计结果驱动下一轮递归精炼的自适应闭环与[recursion]段 Python 代码中的memory_agent_adapt(revised_context, ...)递归调用一一对应。八、实战落地如何接入你的 Agent 系统8.1 直接作为系统提示词注入将 20_templates/PROMPTS/memory.agent.md 全文作为 System Prompt再按[context_schema]填入当前知识库快照JSON即可启动一轮完整的六阶段执行。运行时需注意确保提供session.priority_phases让 Agent 知道本轮聚焦哪些阶段显式声明requested_focus如surfacing用于发现孤立节点避免输出泛化每轮结束后要求 Agent 输出Audit/Version Log形成版本检查点。8.2 以程序化方式组装参考 README 的用法模式20_templates/PROMPTS/README.md 给出了模板的程序化装配思路——用PromptProgram把模板转成可执行的分步程序并用ProtocolShell承载协议定义from templates.prompt_program_template import PromptProgram from templates.field_protocol_shells import ProtocolShell # Load prompt template with open(20_templates/PROMPTS/memory.agent.md, r) as f: template f.read() # Parse context configuration context_config yaml.safe_load(...) # 从 [context_schema] 提取 # Create field protocol protocol ProtocolShell.from_dict(context_config.get(protocol, {})) # Create prompt program with the template program PromptProgram(descriptioncontext_config.get(description, ), templatetemplate) # Execute integrated system result program.execute_with_protocol(protocol, {input: user_query})PromptProgram的实现细节可查看 20_templates/prompt_program_template.py——它支持StepType.INSTRUCTION / CONDITION / LOOP / VARIABLE / FUNCTION / ERROR六种步骤类型并可与NeuralFieldProgram、ProtocolShellProgram组合把六阶段流水线编译为带控制流的分步程序。8.3 与记忆层次实现对接推荐路径若要做成生产级系统建议将模板的六个阶段映射到仓库 00_COURSE/03_context_management/labs/memory_management_lab.py 中的运行时组件模板阶段运行时组件说明ingestHierarchicalMemorySystem.store()按 priority/tags 决定写入工作记忆或长期记忆curateWorkingMemory.cleanup()/LongTermMemory.cleanup()低优先级清理、30 天过期策略semantic_linkLongTermMemory.tag_index 自定义links索引用标签索引近似链接密度contextual_retrieveHierarchicalMemorySystem.search()工作记忆优先 长期记忆补足 可选外部检索recursive_refineHierarchicalMemorySystem.optimize()优先级衰减、降级/晋升、统计输出audit_versionContextWindowManager.assembly_history 自定义版本表记录每次装配与变更其中optimize()返回的统计字典working_memory_freed、long_term_memory_freed、demotions等可直接作为审计日志的量化佐证。8.4 版本迭代建议模板自身带last_updated: 2025-07-08与agent_protocol_version: 1.0.0说明其维护者遵循版本化管理。仓库中还存在姊妹模板 reconstruction.memory.agent.md碎片化存储 上下文驱动装配 AI 缺口填充的重建式记忆二者可按场景选型需要严格溯源与团队协作选/memory.agent节点—链接—审计模型需要对话式动态记忆选/reconstruction.memory.agent碎片—共振—重建模型两者可组合以/memory.agent管理权威知识库以重建式模板处理会话级动态记忆形成权威层 动态层的双层记忆架构。九、总结与使用边界/memory.agent的价值在于把知识库管理从模糊的提示词要求转译为一套可解析的 Schema 可执行的阶段契约 可审计的输出格式 可终止的递归机制。它适合以下场景Agent 长期维护团队/组织知识库需要变更留痕需要语义链接与孤立节点发现的知识图谱运维需要版本检查点与贡献者追溯的合规场景需要随上下文持续重分类的自适应知识系统。使用边界以当前仓库为准模板中ingest_nodes、run_phase等函数为骨架伪代码需结合 00_COURSE/03_context_management/labs/memory_management_lab.py 等实现补全运行时逻辑max_depth6的递归上限需结合上下文窗口参考 40_reference/token_budgeting.md谨慎设定模板面向 GPT-4o、Claude 等主流模型与 Agent 框架跨模型迁移时建议先用[examples]中的表格格式做输出格式校验。从 00_COURSE/03_context_management/02_memory_hierarchies.md 的视角看/memory.agent是把记忆层次从存储架构上升为运维协议的实践范本它证明了记忆管理不只是数据结构问题更是编排、审计与持续精炼的协议问题。赞分享文档教程知识库人工智能提示工程【免费下载链接】Context-EngineeringContext engineering is the delicate art and science of filling the context window with just the right information for the next step. — Andrej Karpathy. A frontier, first-principles handbook inspired by Karpathy and 3Blue1Brown for moving beyond prompt engineering to the wider discipline of context design, orchestration, and optimization.项目地址https://gitcode.com/gh_mirrors/co/Context-Engineering点击查看免费下载相关推荐Context Engineering 实战用 /experiment.agent 系统提示模板构建可审计、可递归的实验设计 AgentContext Engineering 实战用 /experiment.agent 系统提示模板构建可审计、可递归的实验设计 Agent 本文以 Contex文档教程知识库人工智能提示工程Context-Engineering 的 /optimize.agent构建可审计、可递归改进的 Agent 优化系统提示词Context Engineering 的 /optimize.agent构建可审计、可递归改进的 Agent 优化系统提示词 导读本文围绕 Context文档教程知识库人工智能提示工程Context Engineering 实战基于 /research.agent 构建可审计、可组合、递归自改进的研究 Agent 系统提示词Context Engineering 实战基于 /research.agent 构建可审计、可组合、递归自改进的研究 Agent 系统提示词 本文以 Con文档教程知识库人工智能提示工程上一篇Fang worker池配置终极指南并发控制与性能优化技巧下一篇如何使用Projeny5分钟快速上手Unity项目管理神器创作声明：本文部分内容由AI辅助生成（AIGC），仅供参考 Context-Engineering 记忆 Agent 系统提示词模板实战：构建可审计的自适应知识库管理协议
文档教程知识库人工智能提示工程【免费下载链接】Context-EngineeringContext engineering is the delicate art and science of filling the context window with just the right information for the next step. — Andrej Karpathy. A frontier, first-principles handbook inspired by Karpathy and 3Blue1Brown for moving beyond prompt engineering to the wider discipline of context design, orchestration, and optimization.项目地址https://gitcode.com/gh_mirrors/co/Context-Engineering点击查看免费下载导读本文以仓库 20_templates/PROMPTS/memory.agent.md 为核心系统讲解 Context-Engineering 项目中/memory.agent这一套模块化、可递归、可审计的系统提示词System Prompt模板它把知识库与组织记忆的管理拆解为**摄取ingest、策展curate、语义链接link、上下文检索retrieve、递归精炼refine、审计版本audit/version**六个可执行阶段并配套 JSON 上下文 Schema、YAML 工作流、Python 递归适配骨架与完整的 Markdown 输出示例。读完本文你将掌握如何把该模板直接套用于 Agent 化/人工协作的知识库运维场景并能结合仓库中的记忆层次管理实现00_COURSE/03_context_management/labs/memory_management_lab.py落地同构的检索、评分与版本化逻辑。一、模板定位从提示词到记忆管理协议/memory.agent是一份面向自适应知识库/组织记忆管理的模态化 Markdown 系统提示词。它与仓库内其他.agent模板如 research.agent.md、pipeline.agent.md同属 20_templates/PROMPTS/README.md 中定义的任务特定模板 场域操作模板分类——既承担长期记忆管理的任务目标又实现了残留追踪、压缩技术、检索优化的场域概念。模板头部[meta]段声明了其协议元信息{ agent_protocol_version: 1.0.0, prompt_style: multimodal-markdown, intended_runtime: [OpenAI GPT-4o, Anthropic Claude, Agentic System], schema_compatibility: [json, yaml, markdown, python, shell], maintainers: [Recursive Agent Field], audit_log: true, last_updated: 2025-07-08, prompt_goal: Provide a modular, recursive, and auditable prompt template for agentic/human management of knowledge bases or organizational memory—enabling ingestion, semantic linking, adaptive retrieval, recursive categorization, and transparent audit/versioning. }关键设计意图有三点多模态兼容schema_compatibility声明 JSON/YAML/Markdown/Python/Shell 均可承载该模板的数据结构运行时覆盖主流多模态模型与 Agent 框架可审计audit_log: true意味着每一次变更合并、删除、重新分类、链接更新都必须留存贡献者与时间戳递归自适应maintainers标注为 Recursive Agent Field对应下文[recursion]段中的递归精炼机制。这与仓库的Software 3.0理念一脉相承在 00_COURSE/03_context_management/00_overview.md 中Prompt 模板被视为通信层而协议Protocol被视为编排层/memory.agent恰好把两者合成为一个完整的记忆管理编排器。二、模板结构与 ASCII 总览模板正文采用[section]分块组织文件树如下/memory.agent.system.prompt.md ├── [meta] # JSON: protocol version, audit, runtime ├── [ascii_diagrams] # File tree, KB graph, workflow diagrams ├── [context_schema] # JSON: knowledge node, ingestion, session ├── [workflow] # YAML: KB phases and retrieval logic ├── [recursion] # Python: recursive surfacing/categorization ├── [instructions] # Markdown: behavioral rules, DO NOTs ├── [examples] # Markdown: KB entries, curation, logs2.1 知识库的场域结构模板用 ASCII 图描述知识库的拓扑——知识节点经过语义链接形成网络再服务于上下文检索检索结果进入递归浮现/分类最终全部操作落入审计日志[Knowledge Nodes] / | \ v v v [Semantic Linkage] \ | / v v v [Contextual Retrieval] | [Recursive Surfacing/Categorization] | [Audit Log]这条链路与仓库中的记忆层次思想一致在 00_COURSE/03_context_management/02_memory_hierarchies.md 中记忆被组织为即时上下文 → 工作记忆 → 短期存储 → 长期存储 → 归档的多级金字塔而/memory.agent的节点—链接—检索—审计结构正是把该层级思想抽象为知识图谱的运维流程。2.2 六阶段流水线[ingest] | [curate] | [link] | [retrieve] | [refine/recategorize] | [audit/version]这是整个模板的主循环摄取之后必策展策展之后建链接随后按上下文检索检索结果触发递归精炼最后以审计/版本化收尾并进入下一轮。仓库中 00_foundations/03_cells_memory.md 讨论的记忆 Token 预算问题窗口化、摘要化、键值存储、优先级裁剪、语义分块在这里被落实为可操作的六阶段协议。三、上下文 Schema知识节点、摄取记录与会话控制[context_schema]定义了 Agent 解析输入的标准数据结构分为三大块{ knowledge_base: { name: string, domain: string (company, research, ops, product, etc.), nodes: [ { id: string, title: string, content: string, type: string (doc, meeting, insight, spec, etc.), created: timestamp, tags: [string], links: [node_id] } ], link_types: [reference, dependency, related, contradicts, expands, deprecated] }, ingestion: { source: string (doc, chat, code, meeting, email, etc.), method: string (upload, scrape, API, manual, etc.), contributor: string, ingest_time: timestamp }, session: { goal: string, special_instructions: string, priority_phases: [ ingest, curate, semantic_link, contextual_retrieve, recursive_refine, audit_version ], requested_focus: string (discovery, surfacing, onboarding, explainability, etc.) } }3.1 字段语义与取值范围knowledge_base.nodes[].type节点类型枚举doc / meeting / insight / spec可扩展links数组通过node_id建立引用关系。link_types六种语义链接类型——reference引用、dependency依赖、related相关、contradicts矛盾、expands扩展、deprecated已弃用。这是知识图谱的关系谓词集决定了检索与精炼阶段如何理解节点间联系。ingestion.method摄取方式枚举upload / scrape / API / manual每条记录必须携带contributor与ingest_time这是审计可追溯性的数据基础。session.priority_phases允许调用方指定本轮优先执行的阶段子集requested_focus支持discovery发现、surfacing浮现、onboarding入职引导、explainability可解释性等聚焦模式。3.2 与仓库实现的对应关系该 Schema 的节点 标签 优先级 时间戳设计与仓库中的MemoryEntry数据结构高度同构见 00_COURSE/03_context_management/labs/memory_management_lab.pydataclass class MemoryEntry: content: str timestamp: datetime access_count: int 0 last_accessed: Optional[datetime] None priority: float 1.0 size_bytes: int 0 tags: List[str] None两者的差异在于/memory.agent的 Schema 面向 LLM 提示词输入强调可解析的 JSON 语义而 lab 中的MemoryEntry面向运行时实现携带访问计数、优先级衰减、字节大小等工程字段。在实战中可把 Schema 中nodes[].id映射为MemoryEntry的keytags直接复用tag_index做标签检索形成提示词协议 → 运行时记忆层的无缝衔接。四、工作流六个阶段的 YAML 定义[workflow]段用 YAML 形式定义了每个阶段的描述description与输出output这是直接注入提示词的执行契约phases: - ingest: description: | Ingest new knowledge nodes (docs, chats, data) into KB. Record metadata (source, contributor, tags, time). output: - Ingestion table: node, type, tags, source, contributor, timestamp. - curate: description: | Review and clean ingested nodes. Remove duplicates, flag noise, update tags/types. output: - Curation table: node, action (keep/merge/delete), rationale. - semantic_link: description: | Create and update semantic links between nodes. Specify link types (reference, expands, etc.), surface isolated or orphaned nodes. output: - Link map/table: source, target, type, reason. - contextual_retrieve: description: | Retrieve and present nodes relevant to a user’s query/context. Use semantic/contextual cues (tags, recency, link density, etc.). output: - Retrieval table: query/context, retrieved nodes, method, confidence. - recursive_refine: description: | Surface and recategorize nodes/links as new context or queries arise. Adapt taxonomies/tags/relations; propose merges/splits or archive deprecated content. output: - Revision log: phase, change, rationale, timestamp. - audit_version: description: | Log all changes, merges, deletions, recategorizations, and link updates with contributor and timestamp. Surface version checkpoints. output: - Audit/version log: action, node/link, contributor, timestamp, version.4.1 各阶段要点解析阶段核心动作输出产物关键验收点ingest摄取文档/对话/数据记录来源、贡献者、标签、时间摄取表元数据完整性curate去重、标记噪声、更新标签与类型策展表keep/merge/delete rationale每个动作都有理由semantic_link建立/更新语义链接暴露孤立orphaned节点链接映射表链接类型规范、无孤立节点contextual_retrieve依据标签、时效、链接密度等线索检索检索表query nodes method confidence方法透明、置信度标注recursive_refine随新上下文浮现而重新分类建议合并/拆分/归档修订日志变更可回放audit_version记录全部变更并给出版本检查点审计/版本日志贡献者 时间戳 版本号其中contextual_retrieve强调的语义/上下文线索tags、recency、link density与仓库中LongTermMemory.search的评分公式直接对应见 00_COURSE/03_context_management/labs/memory_management_lab.py# Content matching if query_lower in entry.content.lower(): score 0.5 # Tag matching tag_overlap len(query_tags.intersection(set(entry.tags))) if tag_overlap 0: score 0.3 * (tag_overlap / len(query_tags)) # Importance and recency score entry.compute_score() * 0.2从源码结构可以推断tag link检索正是模板方法method列可填写的实现选项之一若接入向量检索则可把 lab 中compute_score里的TODO: semantic similarity补全形成标签召回 语义排序的两级检索。五、递归机制自适应的记忆精炼骨架[recursion]段给出了核心 Python 骨架它定义了整个 Agent 的自适应上限与递归终止条件def memory_agent_adapt(context, stateNone, audit_logNone, depth0, max_depth6): context: dict from context schema state: dict of phase outputs audit_log: list of revision/version entries depth: recursion count max_depth: adaptation limit if state is None: state {} if audit_log is None: audit_log [] # Ingest and curate first state[ingest] ingest_nodes(context, state.get(ingest, {})) state[curate] curate_nodes(context, state.get(curate, {})) # Phased KB operations for phase in [semantic_link, contextual_retrieve, recursive_refine, audit_version]: state[phase] run_phase(phase, context, state) # Recursive surfacing/refinement if depth max_depth and needs_refinement(state): revised_context, reason query_for_refinement(context, state) audit_log.append({revision: phase, reason: reason, timestamp: get_time()}) return memory_agent_adapt(revised_context, state, audit_log, depth 1, max_depth) else: state[audit_log] audit_log return state5.1 设计要点固定入口ingest与curate永远先执行保证后续阶段的输入质量阶段顺序不可变semantic_link → contextual_retrieve → recursive_refine → audit_version严格串行符合先建链接、再检索、后精炼、最后审计的依赖关系递归驱动needs_refinement(state)决定是否进入下一轮每次递归都会在audit_log追加{revision, reason, timestamp}形成修订轨迹深度上限max_depth6防止无限循环。可根据上下文预算调整该值——仓库 40_reference/token_budgeting.md 讨论了 Token 预算约束递归每增加一轮都会消耗额外上下文因此max_depth应结合上下文窗口大小设定。5.2 递归精炼的工程价值递归意味着知识库不是一次性建好而是随新上下文/新查询持续演化分类法taxonomy会变化、标签会更新、关系会重连。这与仓库中HierarchicalMemorySystem.optimize()的优先级衰减 降级/清理循环见 00_COURSE/03_context_management/labs/memory_management_lab.py同构——两者都通过周期性的再评估保持记忆系统的时效性与信噪比。六、行为指令Instructions 与 DO NOT 约束[instructions]段是注入给 LLM 的角色规约逐条如下You are a /memory.agent. You: - Parse all KB, ingestion, and session context from the schema. - Proceed: ingest, curate, semantic link, contextual retrieve, recursive refine, audit/version. - Ask clarifying questions for ambiguous/missing info. - Output all results in labeled Markdown: tables, lists, diagrams. - DO NOT ingest low-signal/noisy, duplicate, or deprecated content without review. - DO NOT skip curation or ignore isolated nodes. - DO NOT break semantic/contextual integrity of links/tags. - Always log rationale and contributors for changes. - Surface version checkpoints after major changes. - Support onboarding with workflow diagrams and file tree. - Close each cycle with audit/version log and summary of open questions.6.1 约束的工程动机DO NOT ingest low-signal/noisy...对应策展阶段的信噪比控制。仓库的WorkingMemory.cleanup()以priority 0.1为清理阈值见 00_COURSE/03_context_management/labs/memory_management_lab.py提示词层面先拦截噪声可减少运行时清理压力DO NOT skip curation or ignore isolated nodes孤立节点是知识图谱的断链信号semantic_link阶段必须显式暴露它们Always log rationale and contributors审计可追溯性的提示词级保障与[workflow]中audit_version的输出契约闭环Surface version checkpoints after major changes对应版本检查点机制是回滚与团队协作的基础。6.2 与协议壳的对照仓库 60_protocols/shells/memory.reconstruction.attractor.shell.md 提供了记忆重建的场域协议壳fragment 扫描 → 共振激活 → 吸引子激励 → 场动力学 → 模式提取 → 缺口识别 → 推理填充 → 一致性验证 → 碎片适配 → 记忆巩固。/memory.agent走的是显式知识图谱路线节点/链接/审计而 attractor shell 走的是神经场重建路线碎片/共振/缺口填充两者互补前者适合需要严格审计的企业知识库后者适合对话式 Agent 的动态记忆。若需理解更深层的场域动力学背景可参考 40_reference/attractor_dynamics.md 与 40_reference/retrieval_indexing.md。七、完整示例从摄取到版本化的输出模板[examples]段提供了可直接复制的输出格式样板覆盖全部六个阶段。这些表格是 Agent 的标准输出协议确保结果可被下游系统解析与审计。7.1 摄取IngestionNode IDTitleTypeTagsSourceContributorTimeN001Q2 Board Recapmeetingops, stratZoomC. Rivera2025-07-08 13:00ZN002API Spec v2docproductDriveK. Chen2025-07-08 13:05ZN003#launch-feedbackchatlaunch, cxSlackT. Adams2025-07-08 13:10Z要点每条记录必须携带来源Source、贡献者Contributor与时间戳Time三者共同构成审计链的起点。7.2 策展CurationNodeActionRationaleN002KeepUnique, up-to-dateN003MergeSimilar to N004N001KeepCore ops recap要点每个动作Keep/Merge/Delete都必须附带 Rationale这是可解释的清理而非随意删减。7.3 语义链接Semantic LinksSourceTargetLink TypeReasonN002N005expandsSpec builds on N005N001N006referenceMeeting covers roadmap要点链接类型严格取自 Schema 中link_types枚举Reason 说明链接成立的理由供后续精炼阶段重估。7.4 上下文检索Contextual RetrievalQueryRetrieved NodesMethodConfidenceAPI launch planN002, N005taglink searchHigh要点Method 列必须显式声明检索策略如taglink search、semantic、recencyConfidence 列标注置信度——对应仓库检索评分的透明度要求。7.5 递归精炼日志Recursive Refinement LogPhaseChangeRationaleTimestampCurationArchived N007Obsolete spec2025-07-08 14:00ZLinkingAdded contradicts N004-N009Prevent confusion2025-07-08 14:01Z7.6 审计/版本日志Audit/Version LogActionNode/LinkContributorTimestampVersionMergeN003/N004T. Adams2025-07-08 14:03Zv1.1CheckpointAllSystem2025-07-08 14:05Zv1.17.7 闭环工作流图[Ingest] - [Curate] - [Link] - [Retrieve] - [Refine] - [Audit] | ^ -------------------------------------------注意图中从Audit回到Ingest/Retrieve的反向箭头——它直观表达了审计结果驱动下一轮递归精炼的自适应闭环与[recursion]段 Python 代码中的memory_agent_adapt(revised_context, ...)递归调用一一对应。八、实战落地如何接入你的 Agent 系统8.1 直接作为系统提示词注入将 20_templates/PROMPTS/memory.agent.md 全文作为 System Prompt再按[context_schema]填入当前知识库快照JSON即可启动一轮完整的六阶段执行。运行时需注意确保提供session.priority_phases让 Agent 知道本轮聚焦哪些阶段显式声明requested_focus如surfacing用于发现孤立节点避免输出泛化每轮结束后要求 Agent 输出Audit/Version Log形成版本检查点。8.2 以程序化方式组装参考 README 的用法模式20_templates/PROMPTS/README.md 给出了模板的程序化装配思路——用PromptProgram把模板转成可执行的分步程序并用ProtocolShell承载协议定义from templates.prompt_program_template import PromptProgram from templates.field_protocol_shells import ProtocolShell # Load prompt template with open(20_templates/PROMPTS/memory.agent.md, r) as f: template f.read() # Parse context configuration context_config yaml.safe_load(...) # 从 [context_schema] 提取 # Create field protocol protocol ProtocolShell.from_dict(context_config.get(protocol, {})) # Create prompt program with the template program PromptProgram(descriptioncontext_config.get(description, ), templatetemplate) # Execute integrated system result program.execute_with_protocol(protocol, {input: user_query})PromptProgram的实现细节可查看 20_templates/prompt_program_template.py——它支持StepType.INSTRUCTION / CONDITION / LOOP / VARIABLE / FUNCTION / ERROR六种步骤类型并可与NeuralFieldProgram、ProtocolShellProgram组合把六阶段流水线编译为带控制流的分步程序。8.3 与记忆层次实现对接推荐路径若要做成生产级系统建议将模板的六个阶段映射到仓库 00_COURSE/03_context_management/labs/memory_management_lab.py 中的运行时组件模板阶段运行时组件说明ingestHierarchicalMemorySystem.store()按 priority/tags 决定写入工作记忆或长期记忆curateWorkingMemory.cleanup()/LongTermMemory.cleanup()低优先级清理、30 天过期策略semantic_linkLongTermMemory.tag_index 自定义links索引用标签索引近似链接密度contextual_retrieveHierarchicalMemorySystem.search()工作记忆优先 长期记忆补足 可选外部检索recursive_refineHierarchicalMemorySystem.optimize()优先级衰减、降级/晋升、统计输出audit_versionContextWindowManager.assembly_history 自定义版本表记录每次装配与变更其中optimize()返回的统计字典working_memory_freed、long_term_memory_freed、demotions等可直接作为审计日志的量化佐证。8.4 版本迭代建议模板自身带last_updated: 2025-07-08与agent_protocol_version: 1.0.0说明其维护者遵循版本化管理。仓库中还存在姊妹模板 reconstruction.memory.agent.md碎片化存储 上下文驱动装配 AI 缺口填充的重建式记忆二者可按场景选型需要严格溯源与团队协作选/memory.agent节点—链接—审计模型需要对话式动态记忆选/reconstruction.memory.agent碎片—共振—重建模型两者可组合以/memory.agent管理权威知识库以重建式模板处理会话级动态记忆形成权威层 动态层的双层记忆架构。九、总结与使用边界/memory.agent的价值在于把知识库管理从模糊的提示词要求转译为一套可解析的 Schema 可执行的阶段契约 可审计的输出格式 可终止的递归机制。它适合以下场景Agent 长期维护团队/组织知识库需要变更留痕需要语义链接与孤立节点发现的知识图谱运维需要版本检查点与贡献者追溯的合规场景需要随上下文持续重分类的自适应知识系统。使用边界以当前仓库为准模板中ingest_nodes、run_phase等函数为骨架伪代码需结合 00_COURSE/03_context_management/labs/memory_management_lab.py 等实现补全运行时逻辑max_depth6的递归上限需结合上下文窗口参考 40_reference/token_budgeting.md谨慎设定模板面向 GPT-4o、Claude 等主流模型与 Agent 框架跨模型迁移时建议先用[examples]中的表格格式做输出格式校验。从 00_COURSE/03_context_management/02_memory_hierarchies.md 的视角看/memory.agent是把记忆层次从存储架构上升为运维协议的实践范本它证明了记忆管理不只是数据结构问题更是编排、审计与持续精炼的协议问题。赞分享文档教程知识库人工智能提示工程【免费下载链接】Context-EngineeringContext engineering is the delicate art and science of filling the context window with just the right information for the next step. — Andrej Karpathy. A frontier, first-principles handbook inspired by Karpathy and 3Blue1Brown for moving beyond prompt engineering to the wider discipline of context design, orchestration, and optimization.项目地址https://gitcode.com/gh_mirrors/co/Context-Engineering点击查看免费下载相关推荐Context Engineering 实战用 /experiment.agent 系统提示模板构建可审计、可递归的实验设计 AgentContext Engineering 实战用 /experiment.agent 系统提示模板构建可审计、可递归的实验设计 Agent 本文以 Contex文档教程知识库人工智能提示工程Context-Engineering 的 /optimize.agent构建可审计、可递归改进的 Agent 优化系统提示词Context Engineering 的 /optimize.agent构建可审计、可递归改进的 Agent 优化系统提示词 导读本文围绕 Context文档教程知识库人工智能提示工程Context Engineering 实战基于 /research.agent 构建可审计、可组合、递归自改进的研究 Agent 系统提示词Context Engineering 实战基于 /research.agent 构建可审计、可组合、递归自改进的研究 Agent 系统提示词 本文以 Con文档教程知识库人工智能提示工程上一篇Fang worker池配置终极指南并发控制与性能优化技巧下一篇如何使用Projeny5分钟快速上手Unity项目管理神器创作声明：本文部分内容由AI辅助生成（AIGC），仅供参考