# nlm-memory - Doramagic AI Context Pack > 定位：安装前体验与判断资产。它帮助宿主 AI 有一个好的开始，但不代表已经安装、执行或验证目标项目。 ## 充分原则 - **充分原则，不是压缩原则**：AI Context Pack 应该充分到让宿主 AI 在开工前理解项目价值、能力边界、使用入口、风险和证据来源；它可以分层组织，但不以最短摘要为目标。 - **压缩策略**：只压缩噪声和重复内容，不压缩会影响判断和开工质量的上下文。 ## 给宿主 AI 的使用方式 你正在读取 Doramagic 为 nlm-memory 编译的 AI Context Pack。请把它当作开工前上下文：帮助用户理解适合谁、能做什么、如何开始、哪些必须安装后验证、风险在哪里。不要声称你已经安装、运行或执行了目标项目。 ## Claim 消费规则 - **事实来源**：Repo Evidence + Claim/Evidence Graph；Human Wiki 只提供显著性、术语和叙事结构。 - **事实最低状态**：`supported` - `supported`：可以作为项目事实使用，但回答中必须引用 claim_id 和证据路径。 - `weak`：只能作为低置信度线索，必须要求用户继续核实。 - `inferred`：只能用于风险提示或待确认问题，不能包装成项目事实。 - `unverified`：不得作为事实使用，应明确说证据不足。 - `contradicted`：必须展示冲突来源，不得替用户强行选择一个版本。 ## 它最适合谁 - **正在使用 Claude/Codex/Cursor/Gemini 等宿主 AI 的开发者**：README 或插件配置提到多个宿主 AI。 证据：`README.md` Claim：`clm_0003` supported 0.86 ## 它能做什么 - **多宿主安装与分发**（需要安装后验证）：项目包含插件或 marketplace 配置，说明它面向一个或多个 AI 宿主的安装和分发。 证据：`.agents/plugins/marketplace.json`, `plugin/.codex-plugin/plugin.json` Claim：`clm_0001` supported 0.86 - **命令行启动或安装流程**（需要安装后验证）：项目文档中存在可执行命令，真实使用需要在本地或宿主环境中运行这些命令。 证据：`README.md` Claim：`clm_0002` supported 0.86 ## 怎么开始 - `npm install -g nlm-memory` 证据：`README.md` Claim：`clm_0004` supported 0.86 - `git clone https://github.com/pbmagnet4/nlm-memory` 证据：`README.md` Claim：`clm_0005` supported 0.86 ## 继续前判断卡 - **当前建议**：需要管理员/安全审批 - **为什么**：继续前可能涉及密钥、账号、外部服务或敏感上下文，建议先经过管理员或安全审批。 ### 30 秒判断 - **现在怎么做**：需要管理员/安全审批 - **最小安全下一步**：先跑 Prompt Preview；若涉及凭证或企业环境，先审批再试装 - **先别相信**：角色质量和任务匹配不能直接相信。 - **继续会触碰**：角色选择偏差、命令执行、宿主 AI 配置 ### 现在可以相信 - **适合人群线索：正在使用 Claude/Codex/Cursor/Gemini 等宿主 AI 的开发者**（supported）：有 supported claim 或项目证据支撑，但仍不等于真实安装效果。 证据：`README.md` Claim：`clm_0003` supported 0.86 - **能力存在：多宿主安装与分发**（supported）：可以相信项目包含这类能力线索；是否适合你的具体任务仍要试用或安装后验证。 证据：`.agents/plugins/marketplace.json`, `plugin/.codex-plugin/plugin.json` Claim：`clm_0001` supported 0.86 - **能力存在：命令行启动或安装流程**（supported）：可以相信项目包含这类能力线索；是否适合你的具体任务仍要试用或安装后验证。 证据：`README.md` Claim：`clm_0002` supported 0.86 - **存在 Quick Start / 安装命令线索**（supported）：可以相信项目文档出现过启动或安装入口；不要因此直接在主力环境运行。 证据：`README.md` Claim：`clm_0004` supported 0.86 ### 现在还不能相信 - **角色质量和任务匹配不能直接相信。**（unverified）：角色库证明有很多角色，不证明每个角色都适合你的具体任务，也不证明角色能产生高质量结果。 - **不能把角色文案当成真实执行能力。**（unverified）：安装前只能判断角色描述和任务画像是否匹配，不能证明它能在宿主 AI 里完成任务。 - **真实输出质量不能在安装前相信。**（unverified）：Prompt Preview 只能展示引导方式，不能证明真实项目中的结果质量。 - **宿主 AI 版本兼容性不能在安装前相信。**（unverified）：Claude、Cursor、Codex、Gemini 等宿主加载规则和版本差异必须在真实环境验证。 - **不会污染现有宿主 AI 行为，不能直接相信。**（inferred）：Skill、plugin、AGENTS/CLAUDE/GEMINI 指令可能改变宿主 AI 的默认行为。 证据：`.agents/plugins/marketplace.json`, `plugin/.codex-plugin/plugin.json` - **可安全回滚不能默认相信。**（unverified）：除非项目明确提供卸载和恢复说明，否则必须先在隔离环境验证。 - **真实安装后是否与用户当前宿主 AI 版本兼容？**（unverified）：兼容性只能通过实际宿主环境验证。 证据：`.agents/plugins/marketplace.json`, `plugin/.codex-plugin/plugin.json` - **项目输出质量是否满足用户具体任务？**（unverified）：安装前预览只能展示流程和边界，不能替代真实评测。 ### 继续会触碰什么 - **角色选择偏差**：用户对任务应该由哪个专家角色处理的判断。 原因：选错角色会让 AI 从错误专业视角回答，浪费时间或误导决策。 - **命令执行**：包管理器、网络下载、本地插件目录、项目配置或用户主目录。 原因：运行第一条命令就可能产生环境改动；必须先判断是否值得跑。 证据：`README.md` - **宿主 AI 配置**：Claude/Codex/Cursor/Gemini/OpenCode 等宿主的 plugin、Skill 或规则加载配置。 原因：宿主配置会改变 AI 后续工作方式，可能和用户已有规则冲突。 证据：`.agents/plugins/marketplace.json`, `plugin/.codex-plugin/plugin.json` - **本地环境或项目文件**：安装结果、插件缓存、项目配置或本地依赖目录。 原因：安装前无法证明写入范围和回滚方式，需要隔离验证。 证据：`.agents/plugins/marketplace.json`, `README.md`, `plugin/.codex-plugin/plugin.json` - **环境变量 / API Key**：项目入口文档明确出现 API key、token、secret 或账号凭证配置。 原因：如果真实安装需要凭证，应先使用测试凭证并经过权限/合规判断。 证据：`README.md`, `SECURITY.md`, `docs/hooks.md`, `docs/install-verification.md` 等 - **宿主 AI 上下文**：AI Context Pack、Prompt Preview、Skill 路由、风险规则和项目事实。 原因：导入上下文会影响宿主 AI 后续判断，必须避免把未验证项包装成事实。 ### 最小安全下一步 - **先跑 Prompt Preview**：先用交互式试用验证任务画像和角色匹配，不要先导入整套角色库。（适用：任何项目都适用，尤其是输出质量未知时。） - **只在隔离目录或测试账号试装**：避免安装命令污染主力宿主 AI、真实项目或用户主目录。（适用：存在命令执行、插件配置或本地写入线索时。） - **先备份宿主 AI 配置**：Skill、plugin、规则文件可能改变 Claude/Cursor/Codex 的默认行为。（适用：存在插件 manifest、Skill 或宿主规则入口时。） - **不要使用真实生产凭证**：环境变量/API key 一旦进入宿主或工具链，可能产生账号和合规风险。（适用：出现 API、TOKEN、KEY、SECRET 等环境线索时。） - **安装后只验证一个最小任务**：先验证加载、兼容、输出质量和回滚，再决定是否深用。（适用：准备从试用进入真实工作流时。） ### 退出方式 - **保留安装前状态**：记录原始宿主配置和项目状态，后续才能判断是否可恢复。 - **准备移除宿主 plugin / Skill / 规则入口**：如果试装后行为异常，可以把宿主 AI 恢复到试装前状态。 - **保留原始角色选择记录**：如果输出偏题，可以回到任务画像阶段重新选择角色，而不是继续沿着错误角色推进。 - **记录安装命令和写入路径**：没有明确卸载说明时，至少要知道哪些目录或配置需要手动清理。 - **准备撤销测试 API key 或 token**：测试凭证泄露或误用时，可以快速止损。 - **如果没有回滚路径，不进入主力环境**：不可回滚是继续前阻断项，不应靠信任或运气继续。 ## 哪些只能预览 - 解释项目适合谁和能做什么 - 基于项目文档演示典型对话流程 - 帮助用户判断是否值得安装或继续研究 ## 哪些必须安装后验证 - 真实安装 Skill、插件或 CLI - 执行脚本、修改本地文件或访问外部服务 - 验证真实输出质量、性能和兼容性 ## 边界与风险判断卡 - **把安装前预览误认为真实运行**：用户可能高估项目已经完成的配置、权限和兼容性验证。 处理方式：明确区分 prompt_preview_can_do 与 runtime_required。 Claim：`clm_0006` inferred 0.45 - **宿主 AI 插件或 Skill 规则冲突**：新规则可能改变用户现有宿主 AI 的工作方式。 处理方式：安装前先检查插件 manifest 和 Skill 文件，必要时隔离测试。 证据：`.agents/plugins/marketplace.json`, `plugin/.codex-plugin/plugin.json` Claim：`clm_0007` supported 0.86 - **命令执行会修改本地环境**：安装命令可能写入用户主目录、宿主插件目录或项目配置。 处理方式：先在隔离环境或测试账号中运行。 证据：`README.md` Claim：`clm_0008` supported 0.86 - **待确认**：真实安装后是否与用户当前宿主 AI 版本兼容？。原因：兼容性只能通过实际宿主环境验证。 - **待确认**：项目输出质量是否满足用户具体任务？。原因：安装前预览只能展示流程和边界，不能替代真实评测。 - **待确认**：安装命令是否需要网络、权限或全局写入？。原因：这影响企业环境和个人环境的安装风险。 ## 开工前工作上下文 ### 加载顺序 - 先读取 how_to_use.host_ai_instruction，建立安装前判断资产的边界。 - 读取 claim_graph_summary，确认事实来自 Claim/Evidence Graph，而不是 Human Wiki 叙事。 - 再读取 intended_users、capabilities 和 quick_start_candidates，判断用户是否匹配。 - 需要执行具体任务时，优先查 role_skill_index，再查 evidence_index。 - 遇到真实安装、文件修改、网络访问、性能或兼容性问题时，转入 risk_card 和 boundaries.runtime_required。 ### 任务路由 - **多宿主安装与分发**：先说明这是安装后验证能力，再给出安装前检查清单。 边界：必须真实安装或运行后验证。 证据：`.agents/plugins/marketplace.json`, `plugin/.codex-plugin/plugin.json` Claim：`clm_0001` supported 0.86 - **命令行启动或安装流程**：先说明这是安装后验证能力，再给出安装前检查清单。 边界：必须真实安装或运行后验证。 证据：`README.md` Claim：`clm_0002` supported 0.86 ### 上下文规模 - 文件总数：304 - 重要文件覆盖：40/304 - 证据索引条目：80 - 角色 / Skill 条目：47 ### 证据不足时的处理 - **missing_evidence**：说明证据不足，要求用户提供目标文件、README 段落或安装后验证记录；不要补全事实。 - **out_of_scope_request**：说明该任务超出当前 AI Context Pack 证据范围，并建议用户先查看 Human Manual 或真实安装后验证。 - **runtime_request**：给出安装前检查清单和命令来源，但不要替用户执行命令或声称已执行。 - **source_conflict**：同时展示冲突来源，标记为待核实，不要强行选择一个版本。 ## Prompt Recipes ### 适配判断 - 目标：判断这个项目是否适合用户当前任务。 - 预期输出：适配结论、关键理由、证据引用、安装前可预览内容、必须安装后验证内容、下一步建议。 ```text 请基于 nlm-memory 的 AI Context Pack，先问我 3 个必要问题，然后判断它是否适合我的任务。回答必须包含：适合谁、能做什么、不能做什么、是否值得安装、证据来自哪里。所有项目事实必须引用 evidence_refs、source_paths 或 claim_id。 ``` ### 安装前体验 - 目标：让用户在安装前感受核心工作流，同时避免把预览包装成真实能力或营销承诺。 - 预期输出：一段带边界标签的体验剧本、安装后验证清单和谨慎建议；不含真实运行承诺或强营销表述。 ```text 请把 nlm-memory 当作安装前体验资产，而不是已安装工具或真实运行环境。 请严格输出四段： 1. 先问我 3 个必要问题。 2. 给出一段“体验剧本”：用 [安装前可预览]、[必须安装后验证]、[证据不足] 三种标签展示它可能如何引导工作流。 3. 给出安装后验证清单：列出哪些能力只有真实安装、真实宿主加载、真实项目运行后才能确认。 4. 给出谨慎建议：只能说“值得继续研究/试装”“先补充信息后再判断”或“不建议继续”，不得替项目背书。 硬性边界： - 不要声称已经安装、运行、执行测试、修改文件或产生真实结果。 - 不要写“自动适配”“确保通过”“完美适配”“强烈建议安装”等承诺性表达。 - 如果描述安装后的工作方式，必须使用“如果安装成功且宿主正确加载 Skill，它可能会……”这种条件句。 - 体验剧本只能写成“示例台词/假设流程”：使用“可能会询问/可能会建议/可能会展示”，不要写“已写入、已生成、已通过、正在运行、正在生成”。 - Prompt Preview 不负责给安装命令；如用户准备试装，只能提示先阅读 Quick Start 和 Risk Card，并在隔离环境验证。 - 所有项目事实必须来自 supported claim、evidence_refs 或 source_paths；inferred/unverified 只能作风险或待确认项。 ``` ### 角色 / Skill 选择 - 目标：从项目里的角色或 Skill 中挑选最匹配的资产。 - 预期输出：候选角色或 Skill 列表，每项包含适用场景、证据路径、风险边界和是否需要安装后验证。 ```text 请读取 role_skill_index，根据我的目标任务推荐 3-5 个最相关的角色或 Skill。每个推荐都要说明适用场景、可能输出、风险边界和 evidence_refs。 ``` ### 风险预检 - 目标：安装或引入前识别环境、权限、规则冲突和质量风险。 - 预期输出：环境、权限、依赖、许可、宿主冲突、质量风险和未知项的检查清单。 ```text 请基于 risk_card、boundaries 和 quick_start_candidates，给我一份安装前风险预检清单。不要替我执行命令，只说明我应该检查什么、为什么检查、失败会有什么影响。 ``` ### 宿主 AI 开工指令 - 目标：把项目上下文转成一次对话开始前的宿主 AI 指令。 - 预期输出：一段边界明确、证据引用明确、适合复制给宿主 AI 的开工前指令。 ```text 请基于 nlm-memory 的 AI Context Pack，生成一段我可以粘贴给宿主 AI 的开工前指令。这段指令必须遵守 not_runtime=true，不能声称项目已经安装、运行或产生真实结果。 ``` ## 角色 / Skill 索引 - 共索引 47 个角色 / Skill / 项目文档条目。 - **Install**（project_doc）：Install · Quick Start · Runtimes · How recall works · Signals · MCP · REST API · Digest · Config · Security · vs Alternatives 激活提示：当用户需要理解项目结构、安装方式或边界时参考。 证据：`README.md` - **nlm-memory pi.dev extension**（project_doc）：This directory is the pi.dev distribution surface for nlm-memory. Pi loads TypeScript extensions declared in ~/.pi/agent/settings.json 's packages array; pointing pi at this directory auto-discovers index.js and registers the prompt-recall hook. 激活提示：当用户需要理解项目结构、安装方式或边界时参考。 证据：`nlm/README.md` - **nlm-memory Hermes Agent plugin**（project_doc）：This directory is the NousResearch Hermes Agent https://github.com/NousResearch/hermes-agent plugin distribution surface for nlm-memory. 激活提示：当用户需要理解项目结构、安装方式或边界时参考。 证据：`plugin-hermes-agent/README.md` - **nlm-memory Codex plugin**（project_doc）：This directory is the Codex plugin distribution surface for nlm-memory. The repo root acts as a Codex marketplace codex plugin marketplace add pbmagnet4/nlm-memory , and this directory is the plugin Codex installs. 激活提示：当用户需要理解项目结构、安装方式或边界时参考。 证据：`plugin/README.md` - **UI components**（project_doc）：Canonical components for the NLM UI. Reach for these before writing new JSX or CSS. If a pattern is repeated three times across pages, it belongs here. 激活提示：当用户需要理解项目结构、安装方式或边界时参考。 证据：`src/ui/components/README.md` - **Backups & restore**（project_doc）：The worst case is a bad migration silently corrupting ~/.nlm/canonical.sqlite . Rolling daily snapshots make that recoverable. 激活提示：当用户需要理解项目结构、安装方式或边界时参考。 证据：`docs/backups.md` - **Code-signal producer nlm code-signal**（project_doc）：Code-signal producer nlm code-signal 激活提示：当用户需要理解项目结构、安装方式或边界时参考。 证据：`docs/code-signal.md` - **Classifier Extraction-Quality Eval**（project_doc）：Scores the truth of what a classifier extracts — are its decisions and entities faithful to the transcript, and does it recover the decisions a strong reference model found? This is the first eval in the repo that measures extraction quality rather than structure JSON validity, counts . Extraction quality is upstream of every recall and precision number the daemon reports, so it gets its own measurable score. 激活提示：当用户需要理解项目结构、安装方式或边界时参考。 证据：`docs/eval-classifier.md` - **Evaluating recall quality**（project_doc）：Three surfaces keep recall quality honest: a per-commit CI gate that fails on regressions, an ad-hoc nlm eval readout an operator runs against their own query set, and a weekly scheduled measurement that records trend over time. 激活提示：当用户需要理解项目结构、安装方式或边界时参考。 证据：`docs/eval-recall.md` - **Evaluating the agent self-improvement signals lane**（project_doc）：Evaluating the agent self-improvement signals lane 激活提示：当用户需要理解项目结构、安装方式或边界时参考。 证据：`docs/eval-signals.md` - **Hooks lifecycle**（project_doc）：What every NLM hook does, when it fires, what it logs, and how it stays out of the way when something goes wrong. 激活提示：当用户需要理解项目结构、安装方式或边界时参考。 证据：`docs/hooks.md` - **Install verification nlm verify**（project_doc）：nlm verify is the release gate. Run it before npm publish and after any fresh install to confirm the daemon is wired and recall works end-to-end. It exits non-zero on any failure so it can gate CI. 激活提示：当用户需要理解项目结构、安装方式或边界时参考。 证据：`docs/install-verification.md` - **Recall baseline methodology**（project_doc）：Where the 97.2% R@5 number on the README comes from, what it measures, and how to reproduce or extend it. 激活提示：当用户需要理解项目结构、安装方式或边界时参考。 证据：`docs/methodology-recall-baseline.md` - **Supersedence and the editable timeline**（project_doc）：Supersedence and the editable timeline 激活提示：当用户需要理解项目结构、安装方式或边界时参考。 证据：`docs/supersedence.md` - **candidate-recall vs ranking — residual-miss diagnosis 2026-06**（project_doc）：candidate-recall vs ranking — residual-miss diagnosis 2026-06 激活提示：当用户需要理解项目结构、安装方式或边界时参考。 证据：`docs/methodology/candidate-recall-vs-ranking-2026-06.md` - **re-derivation rate — design**（project_doc）：re derivation rate is NLM's strategic metric — the operator-outcome number that competitors mem0, agentmemory, Letta cannot match because their destructive lifecycle decay, auto-forget erases the data needed to compute it. It is the headline number for Pulse, the cron digest, and any public marketing scorecard. Detection rule, methodology, and a reproducible script live here so the metric is auditable. 激活提示：当用户需要理解项目结构、安装方式或边界时参考。 证据：`docs/methodology/re-derivation-rate.md` - **recall precision useful-hit — design**（project_doc）：recall precision useful-hit — design 激活提示：当用户需要理解项目结构、安装方式或边界时参考。 证据：`docs/methodology/useful-hit-rate.md` - **FTS5 Lexical Recall Upgrade — Implementation Plan**（project_doc）：FTS5 Lexical Recall Upgrade — Implementation Plan 激活提示：当用户需要理解项目结构、安装方式或边界时参考。 证据：`docs/plans/2026-05-20-fts5-lexical-recall.md` - **Recall Daemon Wedge Fix — Implementation Plan**（project_doc）：Recall Daemon Wedge Fix — Implementation Plan 激活提示：当用户需要理解项目结构、安装方式或边界时参考。 证据：`docs/plans/2026-05-20-recall-daemon-wedge-fix.md` - **NLM Auto-Inject Recall Hook — Design**（project_doc）：NLM Auto-Inject Recall Hook — Design 激活提示：当用户需要理解项目结构、安装方式或边界时参考。 证据：`docs/plans/2026-05-20-recall-hook-design.md` - **NLM Auto-Inject Recall Hook — Implementation Plan**（project_doc）：NLM Auto-Inject Recall Hook — Implementation Plan 激活提示：当用户需要理解项目结构、安装方式或边界时参考。 证据：`docs/plans/2026-05-20-recall-hook-implementation.md` - **FactStore + Storage Port Refactor Implementation Plan**（project_doc）：FactStore + Storage Port Refactor Implementation Plan 激活提示：当用户需要理解项目结构、安装方式或边界时参考。 证据：`docs/plans/2026-05-30-factstore-storage-port.md` - **Supersedence split: replaces vs supersedes**（project_doc）：Supersedence split: replaces vs supersedes 激活提示：当用户需要理解项目结构、安装方式或边界时参考。 证据：`docs/plans/2026-06-10-supersedence-split.md` - **Code-exemplar recall: "what code worked for a task like this"**（project_doc）：Code-exemplar recall: "what code worked for a task like this" 激活提示：当用户需要理解项目结构、安装方式或边界时参考。 证据：`docs/plans/2026-06-15-code-exemplar-recall.md` - **NLM desktop product — phased plan**（project_doc）：Product name. NLM Non-Linear Memory . The repo / package / internal modules keep their current nle-memory codename — renaming code is a separate, invasive task that doesn't block the product. User-facing strings window title, installer, landing page, README hero use NLM. 激活提示：当用户需要理解项目结构、安装方式或边界时参考。 证据：`docs/plans/desktop-product.md` - **FactStore Design Plan nle-memory-ts**（project_doc）：FactStore Design Plan nle-memory-ts 激活提示：当用户需要理解项目结构、安装方式或边界时参考。 证据：`docs/plans/factstore-design.md` - **nlm upgrade command Implementation Plan**（project_doc）：nlm upgrade command Implementation Plan 激活提示：当用户需要理解项目结构、安装方式或边界时参考。 证据：`docs/superpowers/plans/2026-06-02-nlm-upgrade-command.md` - **Agent Self-Improvement Signals Implementation Plan**（project_doc）：Agent Self-Improvement Signals Implementation Plan 激活提示：当用户需要理解项目结构、安装方式或边界时参考。 证据：`docs/superpowers/plans/2026-06-09-agent-self-improvement-signals.md` - **NLM Quality: B+ → A Implementation Plan**（project_doc）：NLM Quality: B+ → A Implementation Plan 激活提示：当用户需要理解项目结构、安装方式或边界时参考。 证据：`docs/superpowers/plans/2026-06-10-nlm-quality-a-plan.md` - **Code-exemplar producer — Phase 1 capture Implementation Plan**（project_doc）：Code-exemplar producer — Phase 1 capture Implementation Plan 激活提示：当用户需要理解项目结构、安装方式或边界时参考。 证据：`docs/superpowers/plans/2026-06-19-exemplar-producer-phase1-capture.md` - **Code-exemplar producer — Phase 2 passive recall Implementation Plan**（project_doc）：Code-exemplar producer — Phase 2 passive recall Implementation Plan 激活提示：当用户需要理解项目结构、安装方式或边界时参考。 证据：`docs/superpowers/plans/2026-06-19-exemplar-producer-phase2-passive-recall.md` - **Oversized-Session Classification + Backlog Recovery Implementation Plan**（project_doc）：Oversized-Session Classification + Backlog Recovery Implementation Plan 激活提示：当用户需要理解项目结构、安装方式或边界时参考。 证据：`docs/superpowers/plans/2026-06-20-oversized-session-classification.md` - **Classifier JSON-Resilience Implementation Plan**（project_doc）：Classifier JSON-Resilience Implementation Plan 激活提示：当用户需要理解项目结构、安装方式或边界时参考。 证据：`docs/superpowers/plans/2026-06-21-classifier-json-resilience.md` - **Retrieval Quality T1 Implementation Plan Steps 1-4**（project_doc）：Retrieval Quality T1 Implementation Plan Steps 1-4 激活提示：当用户需要理解项目结构、安装方式或边界时参考。 证据：`docs/superpowers/plans/2026-06-21-retrieval-quality-t1.md` - **Scheduler Oversized-Session Wiring Implementation Plan 341**（project_doc）：Scheduler Oversized-Session Wiring Implementation Plan 341 激活提示：当用户需要理解项目结构、安装方式或边界时参考。 证据：`docs/superpowers/plans/2026-06-21-scheduler-oversized-wiring.md` - **nlm upgrade command — design spec**（project_doc）：The existing upgrade UX requires an end user to know and run: 激活提示：当用户需要理解项目结构、安装方式或边界时参考。 证据：`docs/superpowers/specs/2026-06-02-nlm-upgrade-command-design.md` - **Kickoff prompt — agent self-improvement signals build**（project_doc）：Kickoff prompt — agent self-improvement signals build 激活提示：当用户需要理解项目结构、安装方式或边界时参考。 证据：`docs/superpowers/specs/2026-06-09-agent-self-improvement-kickoff.md` - **Agent self-improvement signals — design spec**（project_doc）：Agent self-improvement signals — design spec 激活提示：当用户需要理解项目结构、安装方式或边界时参考。 证据：`docs/superpowers/specs/2026-06-09-agent-self-improvement-signals.md` - **Code-exemplar producer: passive recall of historically-beneficial choices**（project_doc）：Code-exemplar producer: passive recall of historically-beneficial choices 激活提示：当用户需要理解项目结构、安装方式或边界时参考。 证据：`docs/superpowers/specs/2026-06-19-exemplar-producer-design.md` - **Testing recall after spec F + G changes**（project_doc）：Testing recall after spec F + G changes 激活提示：当用户需要理解项目结构、安装方式或边界时参考。 证据：`docs/testing-recall.md` - **Security Policy**（project_doc）：NLM is in active development. Security fixes are only backported to the latest minor release. 激活提示：当用户需要理解项目结构、安装方式或边界时参考。 证据：`SECURITY.md` - **Classifier head-to-head — DeepSeek V4 Flash vs qwen3:4b-instruct-2507**（project_doc）：Classifier head-to-head — DeepSeek V4 Flash vs qwen3:4b-instruct-2507 激活提示：当用户需要理解项目结构、安装方式或边界时参考。 证据：`reports/classifier-comparison/2026-06-02-deepseek-v4-vs-qwen3.md` - **Classifier Extraction-Quality Eval — 2026-06-11**（project_doc）：Classifier Extraction-Quality Eval — 2026-06-11 激活提示：当用户需要理解项目结构、安装方式或边界时参考。 证据：`reports/classifier-eval/2026-06-11-extraction-quality.md` - **Recall candidate-stream experiment — NLM 307**（project_doc）：Recall candidate-stream experiment — NLM 307 激活提示：当用户需要理解项目结构、安装方式或边界时参考。 证据：`reports/recall-experiments/2026-06-11-candidate-streams.md` - **Recall ranking phase 2 — NLM 308**（project_doc）：Date: 2026-06-11 Goal: Recover the two miss classes 307 left on the table: - Class 1 ranking : gold already in the keyword candidate set at rank 6–8, edged out by marginally-higher generic BM25 hits. - Class 2 query-side : deep paraphrases ranking 18–37. 激活提示：当用户需要理解项目结构、安装方式或边界时参考。 证据：`reports/recall-experiments/2026-06-11-ranking-phase2.md` - **Fact-recall coverage — NLM 277**（project_doc）：Date: 2026-06-16 Goal: Establish the first recall facts quality baseline no fact-recall benchmark existed — the 307/ 308 work was session-recall only , then fix what it surfaced. 激活提示：当用户需要理解项目结构、安装方式或边界时参考。 证据：`reports/recall-experiments/2026-06-16-fact-recall-coverage.md` - **Stranger-recovery simulation — 2026-06-10**（project_doc）：Stranger-recovery simulation — 2026-06-10 激活提示：当用户需要理解项目结构、安装方式或边界时参考。 证据：`reports/stranger-sim/2026-06-10-recovery-simulation.md` ## 证据索引 - 共索引 80 条证据。 - **Install**（documentation）：Install · Quick Start · Runtimes · How recall works · Signals · MCP · REST API · Digest · Config · Security · vs Alternatives 证据：`README.md` - **nlm-memory pi.dev extension**（documentation）：This directory is the pi.dev distribution surface for nlm-memory. Pi loads TypeScript extensions declared in ~/.pi/agent/settings.json 's packages array; pointing pi at this directory auto-discovers index.js and registers the prompt-recall hook. 证据：`nlm/README.md` - **nlm-memory Hermes Agent plugin**（documentation）：This directory is the NousResearch Hermes Agent https://github.com/NousResearch/hermes-agent plugin distribution surface for nlm-memory. 证据：`plugin-hermes-agent/README.md` - **nlm-memory Codex plugin**（documentation）：This directory is the Codex plugin distribution surface for nlm-memory. The repo root acts as a Codex marketplace codex plugin marketplace add pbmagnet4/nlm-memory , and this directory is the plugin Codex installs. 证据：`plugin/README.md` - **UI components**（documentation）：Canonical components for the NLM UI. Reach for these before writing new JSX or CSS. If a pattern is repeated three times across pages, it belongs here. 证据：`src/ui/components/README.md` - **Package**（package_manifest）：{ "name": "nlm-memory-pi", "version": "0.0.0", "private": true, "description": "pi.dev extension entry for nlm-memory prompt-recall hook.", "type": "module", "main": "index.js" } 证据：`nlm/package.json` - **Package**（package_manifest）：{ "name": "nlm-memory", "version": "0.20.0", "description": "Local-first non-linear memory operating system for AI operators.", "type": "module", "license": "Apache-2.0", "engines": { "node": " =22.0.0" }, "repository": { "type": "git", "url": "git+https://github.com/pbmagnet4/nlm-memory.git" }, "homepage": "https://github.com/pbmagnet4/nlm-memory readme", "bugs": { "url": "https://github.com/pbmagnet4/nlm-memory/issues" }, "keywords": "ai", "memory", "mcp", "claude-code", "codex", "hermes", "local-first", "recall", "session-memory" , "files": "dist", "migrations", "plugin", "plugin-hermes-agent", "nlm", "assets", "LICENSE", "README.md" , "bin": { "nlm": "dist/cli/nlm.js" }, "scripts": { "d… 证据：`package.json` - **Marketplace**（structured_config）：{ "name": "nlm-memory", "interface": { "displayName": "nlm-memory" }, "plugins": { "name": "nlm-memory", "source": { "source": "local", "path": "./plugin" }, "policy": { "installation": "AVAILABLE", "authentication": "ON USE" }, "category": "Coding" } } 证据：`.agents/plugins/marketplace.json` - **Plugin**（structured_config）：{ "name": "nlm-memory", "version": "0.20.0", "description": "Local-first persistent memory for AI coding agents. Recalls relevant prior sessions on every prompt; learns which sessions actually get cited. Owner-edited timeline beats append-only memory layers.", "author": { "name": "Edward Chalupa", "url": "https://github.com/pbmagnet4" }, "license": "Apache-2.0", "homepage": "https://github.com/pbmagnet4/nlm-memory", "repository": "https://github.com/pbmagnet4/nlm-memory", "keywords": "memory", "context", "recall", "agent", "rag", "session" , "mcpServers": "./.mcp.json", "hooks": "./hooks/hooks.json" } 证据：`plugin/.codex-plugin/plugin.json` - **License**（source_file）：Apache License Version 2.0, January 2004 http://www.apache.org/licenses/ 证据：`LICENSE` - **Backups & restore**（documentation）：The worst case is a bad migration silently corrupting ~/.nlm/canonical.sqlite . Rolling daily snapshots make that recoverable. 证据：`docs/backups.md` - **Code-signal producer nlm code-signal**（documentation）：Code-signal producer nlm code-signal 证据：`docs/code-signal.md` - **Classifier Extraction-Quality Eval**（documentation）：Scores the truth of what a classifier extracts — are its decisions and entities faithful to the transcript, and does it recover the decisions a strong reference model found? This is the first eval in the repo that measures extraction quality rather than structure JSON validity, counts . Extraction quality is upstream of every recall and precision number the daemon reports, so it gets its own measurable score. 证据：`docs/eval-classifier.md` - **Evaluating recall quality**（documentation）：Three surfaces keep recall quality honest: a per-commit CI gate that fails on regressions, an ad-hoc nlm eval readout an operator runs against their own query set, and a weekly scheduled measurement that records trend over time. 证据：`docs/eval-recall.md` - **Evaluating the agent self-improvement signals lane**（documentation）：Evaluating the agent self-improvement signals lane 证据：`docs/eval-signals.md` - **Hooks lifecycle**（documentation）：What every NLM hook does, when it fires, what it logs, and how it stays out of the way when something goes wrong. 证据：`docs/hooks.md` - **Install verification nlm verify**（documentation）：nlm verify is the release gate. Run it before npm publish and after any fresh install to confirm the daemon is wired and recall works end-to-end. It exits non-zero on any failure so it can gate CI. 证据：`docs/install-verification.md` - **Recall baseline methodology**（documentation）：Where the 97.2% R@5 number on the README comes from, what it measures, and how to reproduce or extend it. 证据：`docs/methodology-recall-baseline.md` - **Supersedence and the editable timeline**（documentation）：Supersedence and the editable timeline 证据：`docs/supersedence.md` - **candidate-recall vs ranking — residual-miss diagnosis 2026-06**（documentation）：candidate-recall vs ranking — residual-miss diagnosis 2026-06 证据：`docs/methodology/candidate-recall-vs-ranking-2026-06.md` - **re-derivation rate — design**（documentation）：re derivation rate is NLM's strategic metric — the operator-outcome number that competitors mem0, agentmemory, Letta cannot match because their destructive lifecycle decay, auto-forget erases the data needed to compute it. It is the headline number for Pulse, the cron digest, and any public marketing scorecard. Detection rule, methodology, and a reproducible script live here so the metric is auditable. 证据：`docs/methodology/re-derivation-rate.md` - **recall precision useful-hit — design**（documentation）：recall precision useful-hit — design 证据：`docs/methodology/useful-hit-rate.md` - **FTS5 Lexical Recall Upgrade — Implementation Plan**（documentation）：FTS5 Lexical Recall Upgrade — Implementation Plan 证据：`docs/plans/2026-05-20-fts5-lexical-recall.md` - **Recall Daemon Wedge Fix — Implementation Plan**（documentation）：Recall Daemon Wedge Fix — Implementation Plan 证据：`docs/plans/2026-05-20-recall-daemon-wedge-fix.md` - **NLM Auto-Inject Recall Hook — Design**（documentation）：NLM Auto-Inject Recall Hook — Design 证据：`docs/plans/2026-05-20-recall-hook-design.md` - **NLM Auto-Inject Recall Hook — Implementation Plan**（documentation）：NLM Auto-Inject Recall Hook — Implementation Plan 证据：`docs/plans/2026-05-20-recall-hook-implementation.md` - **FactStore + Storage Port Refactor Implementation Plan**（documentation）：FactStore + Storage Port Refactor Implementation Plan 证据：`docs/plans/2026-05-30-factstore-storage-port.md` - **Supersedence split: replaces vs supersedes**（documentation）：Supersedence split: replaces vs supersedes 证据：`docs/plans/2026-06-10-supersedence-split.md` - **Code-exemplar recall: "what code worked for a task like this"**（documentation）：Code-exemplar recall: "what code worked for a task like this" 证据：`docs/plans/2026-06-15-code-exemplar-recall.md` - **NLM desktop product — phased plan**（documentation）：Product name. NLM Non-Linear Memory . The repo / package / internal modules keep their current nle-memory codename — renaming code is a separate, invasive task that doesn't block the product. User-facing strings window title, installer, landing page, README hero use NLM. 证据：`docs/plans/desktop-product.md` - **FactStore Design Plan nle-memory-ts**（documentation）：FactStore Design Plan nle-memory-ts 证据：`docs/plans/factstore-design.md` - **nlm upgrade command Implementation Plan**（documentation）：nlm upgrade command Implementation Plan 证据：`docs/superpowers/plans/2026-06-02-nlm-upgrade-command.md` - **Agent Self-Improvement Signals Implementation Plan**（documentation）：Agent Self-Improvement Signals Implementation Plan 证据：`docs/superpowers/plans/2026-06-09-agent-self-improvement-signals.md` - **NLM Quality: B+ → A Implementation Plan**（documentation）：NLM Quality: B+ → A Implementation Plan 证据：`docs/superpowers/plans/2026-06-10-nlm-quality-a-plan.md` - **Code-exemplar producer — Phase 1 capture Implementation Plan**（documentation）：Code-exemplar producer — Phase 1 capture Implementation Plan 证据：`docs/superpowers/plans/2026-06-19-exemplar-producer-phase1-capture.md` - **Code-exemplar producer — Phase 2 passive recall Implementation Plan**（documentation）：Code-exemplar producer — Phase 2 passive recall Implementation Plan 证据：`docs/superpowers/plans/2026-06-19-exemplar-producer-phase2-passive-recall.md` - **Oversized-Session Classification + Backlog Recovery Implementation Plan**（documentation）：Oversized-Session Classification + Backlog Recovery Implementation Plan 证据：`docs/superpowers/plans/2026-06-20-oversized-session-classification.md` - **Classifier JSON-Resilience Implementation Plan**（documentation）：Classifier JSON-Resilience Implementation Plan 证据：`docs/superpowers/plans/2026-06-21-classifier-json-resilience.md` - **Retrieval Quality T1 Implementation Plan Steps 1-4**（documentation）：Retrieval Quality T1 Implementation Plan Steps 1-4 证据：`docs/superpowers/plans/2026-06-21-retrieval-quality-t1.md` - **Scheduler Oversized-Session Wiring Implementation Plan 341**（documentation）：Scheduler Oversized-Session Wiring Implementation Plan 341 证据：`docs/superpowers/plans/2026-06-21-scheduler-oversized-wiring.md` - **nlm upgrade command — design spec**（documentation）：The existing upgrade UX requires an end user to know and run: 证据：`docs/superpowers/specs/2026-06-02-nlm-upgrade-command-design.md` - **Kickoff prompt — agent self-improvement signals build**（documentation）：Kickoff prompt — agent self-improvement signals build 证据：`docs/superpowers/specs/2026-06-09-agent-self-improvement-kickoff.md` - **Agent self-improvement signals — design spec**（documentation）：Agent self-improvement signals — design spec 证据：`docs/superpowers/specs/2026-06-09-agent-self-improvement-signals.md` - **Code-exemplar producer: passive recall of historically-beneficial choices**（documentation）：Code-exemplar producer: passive recall of historically-beneficial choices 证据：`docs/superpowers/specs/2026-06-19-exemplar-producer-design.md` - **Testing recall after spec F + G changes**（documentation）：Testing recall after spec F + G changes 证据：`docs/testing-recall.md` - **Nlm**（source_file）：import { fileURLToPath } from "node:url"; import { dirname, resolve, join } from "node:path"; import { homedir } from "node:os"; import { mkdirSync, readFileSync, writeFileSync, existsSync, rmSync, copyFileSync } from "node:fs"; import { execFileSync, spawn, spawnSync } from "node:child process"; import { Command } from "commander"; import pkg from "../../package.json" with { type: "json" }; import { serve } from "@hono/node-server"; import { StdioServerTransport } from "@modelcontextprotocol/sdk/server/stdio.js"; import { FactRecallService } from "../core/recall-facts/fact-recall-service.js"; import { RecallService } from "../core/recall/recall-service.js"; import { ProviderRegistry } from… 证据：`src/cli/nlm.ts` - **Citation Detect**（source_file）：import type { ToolUseBlock } from "./transcript.js"; ⋮---- export type CitationKind = "tool use" "prose"; export interface CitationDetectInput { readonly responseText: string; readonly toolUses: ReadonlyArray ; readonly surfacedIds: Iterable ; } export interface DetectedCitation { readonly id: string; readonly kind: CitationKind; } export function detectCitations input: CitationDetectInput : DetectedCitation export function detectCitedIds responseText: string, surfacedIds: Iterable , : string function isNlmTool name: string : boolean function isCiteSessionTool name: string : boolean function safeStringify value: unknown : string 证据：`src/core/hook/citation-detect.ts` - **Gate**（source_file）：export type PromptClass = "generative" "evaluate"; ⋮---- export function classifyPrompt prompt: string : PromptClass 证据：`src/core/hook/gate.ts` - **Pointer Block**（source_file）：export interface PointerHit { readonly id: string; readonly label: string; readonly startedAt: string; readonly summary?: string; } export interface PointerFact { readonly subject: string; readonly predicate: string; readonly value: string; readonly corroborationCount: number; } export interface PointerExemplar { readonly outcome: string; readonly lang: string null; readonly repo: string; readonly taskContext: string; } export function formatPointerBlock hits: ReadonlyArray , facts: ReadonlyArray = , exemplars: ReadonlyArray = , : string 证据：`src/core/hook/pointer-block.ts` - **Strip Injected Context**（source_file）：export function stripInjectedContext text: string : string ⋮---- i = footerIdx; // skip header..footer inclusive ⋮---- function findFooter lines: ReadonlyArray , from: number : number / Collapse 3+ consecutive blank lines left by removal down to one. / function collapseBlankRuns lines: ReadonlyArray : string 证据：`src/core/hook/strip-injected-context.ts` - **Transcript**（source_file）：import { existsSync, readFileSync } from "node:fs"; interface ContentBlock { readonly type: string; readonly text?: string; readonly name?: string; readonly input?: unknown; } interface AssistantMessage { readonly content?: ReadonlyArray string; } interface TranscriptLine { readonly type?: string; readonly message?: AssistantMessage; } export interface ToolUseBlock { readonly name: string; readonly input: unknown; } export interface AssistantTurn { readonly text: string; readonly toolUses: ReadonlyArray ; } ⋮---- function parseTurn parsed: TranscriptLine : AssistantTurn null function readLines transcriptPath: string : string null export function readAllAssistantTurns transcriptPath: string,… 证据：`src/core/hook/transcript.ts` - **Provider Models**（source_file）：import type { ProviderKind, ProviderRow } from "./provider-registry.js"; export type FetchImpl = typeof fetch; ⋮---- interface OllamaTagsResponse { readonly models?: ReadonlyArray ; } interface OpenAIModelsResponse { readonly data?: ReadonlyArray ; } export interface ListModelsOptions { readonly apiKey?: string null; readonly fetchImpl?: FetchImpl; readonly timeoutMs?: number; } export async function listModels provider: ProviderRow, opts: ListModelsOptions = {}, : Promise async function fetchOllamaModels baseUrl: string, fetchImpl: FetchImpl, timeoutMs: number, : Promise async function fetchOpenAIModels baseUrl: string, apiKey: string null, fetchImpl: FetchImpl, timeoutMs: number, : Promise 证据：`src/core/providers/provider-models.ts` - **Fact Recall Service**（source_file）：import type { FactStore } from "@ports/fact-store.js"; import type { LLMClient } from "@ports/llm-client.js"; import { LLMUnreachableError } from "@ports/llm-client.js"; import type { Fact, FactHit, FactMatchField, FactRecallQuery, FactRecallResult, RecallMode, } from "@shared/types.js"; import { tokenSet } from "@core/recall/tokenize.js"; ⋮---- function readBoostCap : number ⋮---- export interface FactRecallServiceDeps { readonly factStore: FactStore; readonly llm: LLMClient; } export class FactRecallService ⋮---- constructor private readonly deps: FactRecallServiceDeps async search input: FactRecallQuery : Promise ⋮---- // A query with no signal at all → empty. Either free-text query, or… 证据：`src/core/recall-facts/fact-recall-service.ts` - **Recall Service**（source_file）：import type { CodeEmbedder } from "@ports/code-embedder.js"; import type { CodeExemplarStore } from "@ports/code-exemplar-store.js"; import type { FactStore } from "@ports/fact-store.js"; import type { LLMClient } from "@ports/llm-client.js"; import { LLMUnreachableError } from "@ports/llm-client.js"; import type { KeywordNeighbor, SemanticNeighbor, SessionStore, } from "@ports/session-store.js"; import type { MatchField, RecallHit, RecallMode, RecallQuery, RecallResult, Session, } from "@shared/types.js"; import { applyFilter } from "./filter.js"; import { keywordMatchFields } from "./match-fields.js"; import { detectQueryShape } from "./query-shape.js"; import { recencyMultiplier } from "… 证据：`src/core/recall/recall-service.ts` - **Related Facts**（source_file）：import type { FactStore } from "@ports/fact-store.js"; import type { Fact, RecallHit, RelatedFact } from "@shared/types.js"; ⋮---- export interface PickRelatedFactsOptions { readonly limit?: number; readonly minCorroboration?: number; readonly minConfidence?: number; } function readInt envKey: string, fallback: number : number function readFloat envKey: string, fallback: number : number export async function pickRelatedFacts hits: ReadonlyArray , factStore: FactStore, opts: PickRelatedFactsOptions = {}, : Promise function collectEntities hits: ReadonlyArray : string 证据：`src/core/recall/related-facts.ts` - **Rewrite Prompt**（source_file）：import { LLMUnreachableError, type RewriteResult } from "@ports/llm-client.js"; ⋮---- export function parseRewriteJson raw: string, providerLabel: string : RewriteResult 证据：`src/core/recall/rewrite-prompt.ts` - **Aggregate**（source_file）：import type { Signal } from "@shared/types.js"; export interface FailureMode { readonly repo: string; readonly model: string; readonly kind: string; readonly step: string null; readonly total: number; readonly failures: number; readonly failRate: number; readonly lastTs: string; } export interface AggregateOptions { readonly minFailRate?: number; readonly minSamples?: number; } ⋮---- export function aggregateFailureModes signals: ReadonlyArray , opts: AggregateOptions = {}, : ReadonlyArray ⋮---- type Bucket = { repo: string; model: string; kind: string; step: string null; total: number; failures: number; lastTs: string }; 证据：`src/core/signals/aggregate.ts` - **Pg Code Exemplar Store**（source_file）：import type { Pool } from "pg"; import type { CodeExemplarSearchFilter, CodeExemplarStore, ExemplarVerdictPatch, ExemplarVerdictResult, ExemplarVerdictSource } from "@ports/code-exemplar-store.js"; import type { CodeExemplar, CodeExemplarHit, CodeExemplarInput, CodeExemplarOutcome } from "@shared/types.js"; import { exemplarId } from "./sqlite-code-exemplar-store.js"; ⋮---- interface ExemplarRow { id: string; install scope: string; signal id: string null; session id: string null; repo: string; model: string; lang: string null; task context: string; code: string; code hash: string; outcome: CodeExemplarOutcome; git sha: string null; survived: number null; ts: string; created at: string; reti… 证据：`src/core/storage/pg-code-exemplar-store.ts` - **Sqlite Code Exemplar Store**（source_file）：import { createHash } from "node:crypto"; import type Database from "better-sqlite3"; import type { CodeExemplarSearchFilter, CodeExemplarStore, ExemplarVerdictPatch, ExemplarVerdictResult, ExemplarVerdictSource } from "@ports/code-exemplar-store.js"; import type { CodeExemplar, CodeExemplarHit, CodeExemplarInput, CodeExemplarOutcome } from "@shared/types.js"; ⋮---- type ExemplarRow = { id: string; install scope: string; signal id: string null; session id: string null; repo: string; model: string; lang: string null; task context: string; code: string; code hash: string; outcome: CodeExemplarOutcome; git sha: string null; survived: number null; ts: string; created at: string; retired at: str… 证据：`src/core/storage/sqlite-code-exemplar-store.ts` - **Pagination**（source_file）：interface PaginationProps { page: number; pageSize: number; total: number; pageSizes?: readonly number ; onPageChange: page: number = void; onPageSizeChange: pageSize: number = void; } ⋮----