20:"$Sreact.view_transition" 21:I[9754,["/_next/static/chunks/2t36f5p4pjzlb.js","/_next/static/chunks/1gyqeidjm5eoa.js","/_next/static/chunks/0_fjnxc4tt_l_.js","/_next/static/chunks/1ymt1shyhmu-4.js","/_next/static/chunks/3kv2zf0i7s2wc.js","/_next/static/chunks/1vs_3g_g75zmk.js","/_next/static/chunks/2g4dh802s8oz3.js","/_next/static/chunks/1bc3f0shukf1j.js","/_next/static/chunks/192kgs9v_t14j.js","/_next/static/chunks/0z6pw0v5c_lx5.js","/_next/static/chunks/08r60e6icmtkc.js","/_next/static/chunks/3trx6c0i7_x-3.js","/_next/static/chunks/40sro5bzp9bi_.js","/_next/static/chunks/1wgf3tk38y6cf.js"],"ProjectDetail"] 22:T58fd, agentmemory · 让编码代理记住一切

github.com/rohitg00/agentmemory

让编码代理记住一切，不再重复解释

agentmemory 为 AI 编码代理提供持久记忆，让上次会话里讲过的背景，自动延续到下一次。MCP、hooks、REST API，都能接。

开始使用 npx @agentmemory/agentmemory

首次运行即交互式配置，内存服务器默认在 :3111

此刻，代理在 :3111 处命中了记忆节点——它往回够到过去的会话，而不是从零开始。

一条贯穿所有章节的召回线

01 · 问题的缝隙

每次新会话，都要重新解释一遍

没有记忆的编码代理，每次打开会话都把项目背景当成本次才知道的事。你再次说明架构、偏好和上下文，它也再次现学。这条缝隙，就是重复劳动。

02 · 命中记忆

一次回取，拉回整条过去

agentmemory 把每次会话存下来。当下一个问题到来，代理在 :3111 的内存服务器上命中相关上下文，而不是重新来过。

阅读回取的 R@5 命中率（项目自述基准）

README · 项目声称，非独立评测

03 · 存下并接通

用你已经在用的方式接入

记忆不只是写给单个代理专用，它能经由 MCP、hooks 与 REST API 进入任何工作流，也能直接导入历史会话。

ROUTE

MCP

任何支持 MCP 的客户端都能读写这段持久记忆。

~/.claude/projects

ROUTE

Hooks

在命令行代理的生命周期里接入，把对话写入记忆服务器。

memory server @ :3111

ROUTE

REST API

用自己写的流程调用记忆，不绑定具体某个代理。

import / export session

04 · 一条线，派给所有代理

20+ 代理，读同一份记忆

一条召回线广播到各个代理适配器。Claude Code、Cursor、Codex CLI 都能命中同一批上下文，不问对方是谁。

Claude Code Cursor Codex CLI Gemini CLI Hermes Devin OpenClaw + more

原生技能

17 个原生技能

安装 17 个原生技能，让代理知道该在什么时候伸手去取记忆，而不是每次都要你提醒。

npx skills add rohitg00/agentmemory -y

05 · 真机命令

从一条命令开始

上手只需一行。下面是真实会用到的几条命令与记忆路径。

启动 npx @agentmemory/agentmemory

演示 npx @agentmemory/agentmemory demo

导入 npx @agentmemory/agentmemory import-jsonl ~/.claude/projects

升级 npx @agentmemory/agentmemory upgrade

06 · 项目自述的成效

更少 token，更高命中

agentmemory 主张用持久记忆替代重复解释，从而省下 token。以下数字来自 README 的自述基准，属于项目声称的成效，而非独立测试。

R@5 命中率 · 声称减少的 token 消耗

基准数据来自项目自身（README），使用时请结合自身场景验证。

23:T55cc, agentmemory · Your coding agent remembers everything

github.com/rohitg00/agentmemory

Your coding agent remembers everything. No re-explaining.

agentmemory gives AI coding agents persistent memory, so the context you set up in one session carries into the next — wired through MCP, hooks, and the REST API.

Quick start

npx @agentmemory/agentmemory

First run is interactive setup; the memory server starts on :3111

Right now the agent hits a recall node at :3111 — reaching back for past sessions instead of starting over.

One thread pulled through every section

01 · The re-ask gap

Every new session, you re-explain

A memory-less coding agent treats each session as a fresh start. You re-state the architecture, the preferences, the context — and it relearns them from scratch. That gap is the wasted repetition.

02 · The recall hit

One retrieval pulls back the past

agentmemory stores each session. When the next question comes, the agent hits relevant context on its :3111 memory server instead of beginning again.

R@5 retrieval accuracy (project-stated benchmark)

README · claim, not an independent test

03 · Store & wire it in

Integrate the way you already work

Memory isn't locked to one agent. It flows into any workflow through MCP, hooks, and REST — and can import the sessions you already have.

ROUTE

MCP

Any MCP client can read and write this persistent memory.

~/.claude/projects

ROUTE

Hooks

Hook into a CLI agent's lifecycle and write the conversation to the memory server.

memory server @ :3111

ROUTE

REST API

Drive memory from your own flows, independent of any single agent.

import / export session

04 · One thread, fanned to every agent

20+ agents read the same memory

One recall thread broadcasts to the agent adapters. Claude Code, Cursor, and Codex CLI all hit the same context — no matter who's asking.

Claude Code Cursor Codex CLI Gemini CLI Hermes Devin OpenClaw + more

Native skills

17 native skills

Install 17 native skills so your agent knows when to reach for memory — instead of you reminding it each time.

npx skills add rohitg00/agentmemory -y

05 · Commands that actually run

Start from a single command

Get going with one line. These are the real commands and memory paths you'll reach for.

Start npx @agentmemory/agentmemory

Demo npx @agentmemory/agentmemory demo

Import npx @agentmemory/agentmemory import-jsonl ~/.claude/projects

Upgrade npx @agentmemory/agentmemory upgrade

06 · Stated outcomes

Fewer tokens, sharper hits

agentmemory argues that persistent memory replaces re-explanation and saves tokens. These numbers are the README's own stated benchmarks — claimed outcomes, not independent tests.

R@5 accuracy · claimed token savings

Benchmarks come from the project itself (README); verify against your own setup before relying on them.

a:[["$","$20",null,{"enter":{"secondary-forward":"secondary-forward","secondary-back":"secondary-back-enter","default":"none"},"exit":{"secondary-forward":"secondary-forward","secondary-back":"secondary-back-exit","default":"none"},"update":{"secondary-forward":"secondary-forward","secondary-back":"secondary-back-update","default":"none"},"default":"none","children":["$","$L21","cmt2zh2nx01gf62717xue2pg6:zh-CN:beautified",{"project":{"id":"cmt2zh2nx01gf62717xue2pg6","slug":"rohitg00-agentmemory","name":"agentmemory","description":"为 AI 编码代理提供持久记忆","authorName":"rohitg00","authorVerification":{"verified":false,"verifiedAt":null},"category":"development_tools","subcategory":"data_ai_tools","experienceUrl":"https://agent-memory.dev/","qualityScore":90,"featured":false,"officialSite":null,"views":{"original":{"available":true},"official":{"available":false,"reason":"disabled_by_policy"},"beautified":{"available":true,"reason":null},"defaultView":"beautified"},"beautifiedView":{"revisionId":"cmt2zh2oh01gr62714gzsrddf","kind":"artifact","artifact":{"schemaVersion":"localized-static-html-v4","kind":"static_html","locales":["zh-CN","en"],"defaultLocale":"zh-CN","documents":{"zh-CN":"$22","en":"$23"},"designDirections":{"schemaVersion":"dodot-design-directions-v1","directions":[{"id":"recall-thread","name":"Recall Thread","projectMetaphor":"Memory as a continuous thread pulled forward across sessions — each past exchange stays threaded into the present, so nothing needs re-experiencing. The moment of recall is a single bright node where the agent reaches back instead of starting over.","signatureVisual":"A horizontal braid of session beads that recedes into the background, with one foreground node lit as the instant memory is retrieved; a caption tracks that node back to prior sessions.","heroGeometry":"Horizontally layered ribbons compressing behind a single foreground focus node — a focal plane where memory 'hits' today while past threads fall away in time.","sectionSequence":["the-reask-gap","recall-hit","store-and-hooks","reach-adapters","runCommands","recallBenchmark"],"surfaceModel":"Thread woven on near-black ink; layered ribbon surfaces with fine web-like trace; depth expressed by thread opacity and strand density, not by shadows or cards.","palette":"Ink #0e1113 field, warm thread gold #f0b75f for recall, cool trace #7fd4c8 for persistence, paper #f3efe6 text, faint strand #6b7a85.","typography":"Editorial serif display for the recall node and section numbers; monospaced labels for commands and memory paths; light sans body.","desktopComposition":"Hero: single bright recall node at right center with strands receding leftward across the plate. Sections read as one continuous threaded spine, with the composition animated as strands being pulled into a node at each heading.","mobileTransformation":"The horizontal braid rotates 90° into a vertical stacked thread; ribbons fall in sequence from top to a bottom focus node, preserving the 'pull into node' idea at 390px.","mediaStrategy":"No photography. Derived 'thread-to-node' figure is itself the signature visual. No screenshots (none present). No fake terminal or browser frame.","forbiddenPatterns":["generic slogan over three metric cards","feature-card walls that arbitrary projects could reuse","browser or terminal shell framing","glow/grid/particle field as decorative filler"],"compositionFingerprint":"Two-tone thread braid receding to a single lit recall node, threaded section spine with monospaced memory-path labels."},{"id":"index-fanout","name":"Index Fanout","projectMetaphor":"Memory as a retrieval index behind an addressing core: many agents ask one store, and the answer fans back across all of them. A single core, many unified addresses.","signatureVisual":"A central memory core with radial adapter spokes (one per agent) converging inward; a highlight trace shows a query answered once and fanned back outward.","heroGeometry":"Concentric convergence rings shrinking toward a compact core dot at the left, with adapter spokes reaching outward to the right edge.","sectionSequence":["index-problem","core-addressing","query-once","fan-adapters","probe-skills","reference-benchmark"],"surfaceModel":"Blueprint grid plane over deep field; surfaces are flat engineering plates and ruled alignment marks; depth via concentric ring scale rather than elevation.","palette":"Deep blueprint #0b1230 field, white ink #eef2ff, core amber #f5c65d, adapter teal #67e8d6, faint grid #1e2a55.","typography":"Monospaced technical headings with index notation; condensed sans for adapter tags; tabular figures for retrieval numbers.","desktopComposition":"Hero: concentric rings pin to core at left, spokes fan to a wide adapter band on the right. Sections follow the rings outward as the page 'reads' the index left-to-right.","mobileTransformation":"Radial fan collapses into stacked horizontal tick rows; each tick aligns to a shared left core gutter, folding the rings into a vertical index list.","mediaStrategy":"No photography. The radial core-and-spoke figure is built purely from verified adapter relation (20+ agents) and rendered as a derived diagram, not a product shot.","forbiddenPatterns":["generic slogan over metric cards","feature-card walls with arbitrary copy","terminal or editor shell framing","decorative glow as a substitute for the index idea"],"compositionFingerprint":"Concentric rings converging on a core dot at the left, adapter spokes fanning right, index-notation headings with tabular figures."},{"id":"ledger-replay","name":"Ledger Replay","projectMetaphor":"Memory as a ledger of past sessions that can be skimmed and replayed — each chapter is a page, and replaying an old chapter saves token costs instead of rewriting it.","signatureVisual":"A stacked ledger of session chapters whose top chapter lifts forward; a marginal annotation shows the replayed chapter reclaiming tokens.","heroGeometry":"Vertically stacked chapter spines with one spine pulled toward the viewer at front-center, others receding depth-wise behind it.","sectionSequence":["forget-cost","remember-page","replay-old","economize-tokens","plug-adapters","import-archive"],"surfaceModel":"Warm card-stock ledger pages with stitching and margin rules; surfaces are paper leaves with deckle edge; depth via page overlap rather than shadows.","palette":"Warm paper #f4efe6 field, deep ink #16130e text, ledger bourbon #b4702f, saved-token sage #4e8d78, faint margin #c8bfae.","typography":"Bookish serif headings with chapter numerals; italic margin annotations; monospaced for import commands.","desktopComposition":"Hero: ledger stack at center, top chapter pulled forward as the replay moment, marginal token note beside it. Sections read as chapters opening top-to-bottom.","mobileTransformation":"The ledger stack flattens into a top-aligned accordion of chapter rows; the lift becomes a tap-target that pulls the row forward, preserving the replay gesture at 390px.","mediaStrategy":"No photography. The lifted-chapter figure is a derived composition of the session-replay and import relation (import ~/.claude/projects), not a screenshot.","forbiddenPatterns":["generic slogan over metric cards","feature-card walls with arbitrary copy","terminal or dashboard shell framing","glow/grid field as decoration"],"compositionFingerprint":"Stacked ledger chapters with forward-lifted replay spine, margin token annotations, chapter-numeral serif headings."}],"selectedDirectionId":"recall-thread","selectionRationale":"The recall-thread direction best expresses the project's actual mechanism: an agent reaches back across sessions and pulls prior context forward instead of re-explaining. Its 'bright node where memory is hit' hero maps directly to the verified readme claim of a memory server at `:3111` returning stored context, and its threaded spine supports all evidence sections (the re-ask gap, recall hit, storage via hooks/MCP, broadcast to 20+ adapters, and the self-reported recall/token benchmarks) as one continuous idea rather than stacked marketing blocks. It also departs most clearly from the forbidden defaults and from the V3 seed's dark marketing-hero grammar, so the rebuild yields a genuinely project-specific, evidence-driven visual."},"designBrief":{"schemaVersion":"dodot-design-brief-v1","selectedDirectionId":"recall-thread","projectMetaphor":"Memory as a continuous thread pulled forward across sessions — each past exchange stays threaded into the present, so nothing needs re-experiencing. The moment of recall is a single bright node where the agent reaches back instead of starting over.","signatureVisual":"A horizontal braid of session beads that recedes into the background, with one foreground node lit as the instant memory is retrieved; a caption tracks that node back to prior sessions.","compositionFingerprint":"Two-tone thread braid receding to a single lit recall node, threaded section spine with monospaced memory-path labels.","heroGeometry":"Horizontally layered ribbons compressing behind a single foreground focus node — a focal plane where memory 'hits' today while past threads fall away in time.","sectionSequence":["the-reask-gap","recall-hit","store-and-hooks","reach-adapters","runCommands","recallBenchmark"],"surfaceModel":"Thread woven on near-black ink; layered ribbon surfaces with fine web-like trace; depth expressed by thread opacity and strand density, not by shadows or cards.","palette":"Ink #0e1113 field, warm thread gold #f0b75f for recall, cool trace #7fd4c8 for persistence, paper #f3efe6 text, faint strand #6b7a85.","typography":"Editorial serif display for the recall node and section numbers; monospaced labels for commands and memory paths; light sans body.","desktopComposition":"Hero: single bright recall node at right center with strands receding leftward across the plate. Sections read as one continuous threaded spine, with the composition animated as strands being pulled into a node at each heading.","mobileTransformation":"The horizontal braid rotates 90° into a vertical stacked thread; ribbons fall in sequence from top to a bottom focus node, preserving the 'pull into node' idea at 390px.","mediaStrategy":"No photography. Derived 'thread-to-node' figure is itself the signature visual. No screenshots (none present). No fake terminal or browser frame.","evidencePlan":["Demonstrate the re-ask gap using the overview claims: agents re-explain each session (persistent memory avoids each-session re-explanation).","Show the recall hit via the memory server on `:3111` and the retrieval accuracy claim (R@5 95.2%) from the readme summary.","Ground store-and-hooks in verified integration routes: MCP, hooks, REST API, and import of `~/.claude/projects`.","Ground reach-adapters in the 20+ agent support and 17 native skills claims (Claude Code, Cursor, Codex CLI).","Ground the economy claim (92% fewer tokens) as a project-stated benchmark, labeled as such, not as independent fact.","Use only the single install command `npx @agentmemory/agentmemory` as the real CTA output."],"localeContract":{"zh-CN":"让编码代理记住一切 · 每次会话不再重复解释。表达自然的中文密度，命令与记忆路径保持英文原文。","en":"Your coding agent remembers everything · no re-explaining across sessions. Natural English density, commands and memory paths kept verbatim."},"forbiddenPatterns":["generic slogan over three metric cards","feature-card walls that arbitrary projects could reuse","browser or terminal shell framing","glow/grid/particle field as decorative filler"],"historicalReuse":{"sourceAssetId":"cmsv8evy5006jbnqtd21i8eko","decision":"rebuild","rationale":"The V3 artifact is a dark marketing landing page built on the forbidden defaults (a slogan hero, glow/grid backgrounds, and rectangular feature sections), which does not express the recall-before-prompt mechanism. It is treated strictly as isolated visual history, not project fact or instruction. I rebuild the page's composition entirely around the threaded recall node while carrying forward only a few tone-consistent visual ingredients.","reusedElements":["warm gold accent tonality for emphasis","dark ink field tonality for the surface","condensed uppercase kicker label styling above section titles"],"discardedElements":["marketing slogan hero with radial glow backdrop","rectangular card-grid section grammar","three-metric-card structure","generic feature panel wall","all V3 copy and any unverified numbers presented without source labels"]}},"visualReview":{"schemaVersion":"dodot-automated-design-review-v1","status":"approved","summary":"忠实实现 recall-thread Brief：Hero 为水平编织线程拉向单个点亮的召回节点，纯 monospace+serif 表面、精确调色板、全部数字标注为项目自述基准并附声明，双语结构等价且自然，命令/记忆路径逐字保留，无通用 Hero/三卡/卡片墙/设备壳/装饰渐变，且 390px 手机断点使用独立纵向 hero-svg-mobile（自上而下的线程与串珠落入底部召回节点）真正重构了签名视觉，无假控件。七项检查全部通过，无 blocking issue。","checks":{"projectFit":{"passed":true,"evidence":"Hero 桌面 SVG（hero-svg-desktop，viewBox 1200x760）绘制 5 条双色渐变 strands、6 颗随透明度增大的 session beads（cx 180→760）收敛到一个点亮的金色召回节点（x=920,y=452，r=26 实心+r=46 光晕+r=52 圆环），并有 thread pulled into node 路径及 monospace hero-node-caption 把节点追踪回历史会话，完全对应 brief 的 heroGeometry/desktopComposition/signatureVisual。调色板逐字匹配 brief（ink #0e1113、gold #f0b75f、teal #7fd4c8、paper #f3efe6、strand #6b7a85），serif 章节数字 + monospace 命令/记忆路径标签。章节顺序 the-reask-gap/recall-hit/store-and-hooks/reach-adapters/runCommands/recallBenchmark 与 brief sectionSequence 一致。选中的 selectedDirectionId=recall-thread，无 index-fanout/ledger-replay 方向泄漏。"},"evidenceBound":{"passed":true,"evidence":"全部事实均可在 material.json 中找到依据：95.2% R@5 与 92% token 减少均在页面中标为『项目自述基准/README claim, not an independent test』并附 caveat 声明（非独立事实）；:3111 内存服务器、MCP/hooks/REST、~/.claude/projects 导入、20+ 适配器、17 原生技能、安装命令 npx @agentmemory/agentmemory 均对应提取事实与 readme.commands；0.9.29/Cursor 插件、Devin/DeepSeek harness 等未虚构。reach-adapters 标签（Claude Code/Cursor/Codex CLI/Gemini CLI/Hermes/Devin/OpenClaw）皆有 README/lifecycle 依据。"},"signatureVisual":{"passed":true,"evidence":"thread-to-node 图在源码中真实存在且专属：桌面 SVG 用 5 条渐变 strands + 6 颗递增串珠 + 高亮召回节点 + pull-in 路径组成『编织线程拉向单个点亮节点』；底部 recall-hit 与 recallBenchmark 的 SVG 图继续沿用 hit-node 与 re-explaining-falls-away 曲线，图中含 :3111 命中节点标注。该图形是项目机制的真实表达，并非装饰性渐变/粒子填充。"},"mobileTransformation":{"passed":true,"evidence":"Brief 要求手机端把水平编织旋转 90° 成纵向、自上而下落入底部焦点节点。源码实现了独立 hero-svg-mobile（viewBox 390x1080），在 @media(max-width:480px) 下 hero-svg-desktop display:none 而 hero-svg-mobile display:block：5 条 strands 从顶部向下流动、6 颗串珠 cy 140→720 递增，终点为底部点亮召回节点（cx=198,cy=880，r=30 实心+r=52 光晕+r=60 圆环）并有 thread 向下拉入底节点的路径。@media(max-width:860px) 使 .hit/.adapt/.bench/.routes/.cmd-row 塌缩为单列，480px 下 .hero-node-caption 移到底部，保留了『拉入节点』的意图，满足 390px 重构要求。"},"localeParity":{"passed":true,"evidence":"zh-CN 与 en 页面结构完全一致：相同的 7 个 data-dodot-section、相同的 6 个数字 kicker、相同命令与 SVG 图、相同统计数字。两种语言文案自然（zh-CN 全中文，latin 仅用于命令/路径/语法；en 全英文），命令与记忆路径在双语中逐字保留，符合 localeContract。zh-CN 无中文码点混入命令，en 无英文出现于 label 外的中文。无任一章节/事实/图偏移。"},"antiTemplate":{"passed":true,"evidence":"未回退到通用 Hero/三卡/卡片墙/设备壳/装饰网格：统计以带来源标注的项目自述数字呈现而非 slogan-over-3-metric-cards；store-and-hooks 的三个 ROUTE 是扎根于真实集成路径（MCP/hooks/REST）的细线列表而非可复用浮空卡片墙；无浏览器/终端外壳框（V3 的设备外壳与 radial-gradient 背景已被丢弃，body 用纯色 --bg）；无发光/网格/粒子作为装饰填充。排版与组成由线程脊柱与召回图驱动。"},"noFakeUi":{"passed":true,"evidence":"每个命令均可追溯 readme.commands：npx @agentmemory/agentmemory、demo、import-jsonl ~/.claude/projects、upgrade、以及 npx skills add rohitg00/agentmemory -y 全部真实存在。无假终端或浏览器 chrome。qa-report.json 对每个 locale x viewport 均记录 fakeControlCount 0、brokenImageCount 0、missingAltCount 0、emptyLinkCount 0、无 consoleErrors、无 networkRequests。"}},"issues":[]},"repairEvidence":[{"attempt":1,"before":{"qaPassed":true,"designReviewApproved":false,"qaEvidence":{"zh-CN":[{"viewport":"desktop","passed":true,"localePassed":true,"metrics":{"language":"zh-CN","h1Count":1,"h2Count":6,"sectionCount":7,"textLength":1548,"cjkCount":641,"latinLetterCount":536,"scrollWidth":1280,"clientWidth":1280,"scrollHeight":4690,"h1Top":236.3125,"bottomInset":266.921875,"brokenImageCount":0,"missingAltCount":0,"fakeControlCount":0,"emptyLinkCount":0},"screenshotHashes":{"top":"5985320d99c6bc721729d500bc5d0efd4907c6cabb36ecd175b560c51e737b68","middle":"65070ceb39861f4e935eab060d4f9c10255597de42ca029bf5640b2a72419231","bottom":"babb46ddefc2ceadad3c0023980314a721d62667c6a452d0ac139063cc12cf31"},"consoleErrors":[],"networkRequests":[]},{"viewport":"mobile","passed":true,"localePassed":true,"metrics":{"language":"zh-CN","h1Count":1,"h2Count":6,"sectionCount":7,"textLength":1548,"cjkCount":641,"latinLetterCount":536,"scrollWidth":390,"clientWidth":390,"scrollHeight":5862,"h1Top":148.859375,"bottomInset":374.578125,"brokenImageCount":0,"missingAltCount":0,"fakeControlCount":0,"emptyLinkCount":0},"screenshotHashes":{"top":"9385e9006b2e8f7ade9dc5ec06492535f77d7d2cc0d884496b2dbc31519e018e","middle":"2a9ec69d302277624050a72734ba0bacbdc7bbed970c6f8747f1f9727e633a68","bottom":"a2ce9ab4256b6b2871810d7fe68f3758d0c176ee76ce4d615e78480efe039ed1"},"consoleErrors":[],"networkRequests":[]}],"en":[{"viewport":"desktop","passed":true,"localePassed":true,"metrics":{"language":"en","h1Count":1,"h2Count":6,"sectionCount":7,"textLength":3000,"cjkCount":0,"latinLetterCount":2312,"scrollWidth":1280,"clientWidth":1280,"scrollHeight":4933,"h1Top":113.828125,"bottomInset":266.890625,"brokenImageCount":0,"missingAltCount":0,"fakeControlCount":0,"emptyLinkCount":0},"screenshotHashes":{"top":"b4ec03de74692489646bc519e23c43bad436beba109a2b465e3c9962b792d1d2","middle":"8942ec4fe05ff192f2cf5dcc450d0b60cbdf27fe34dc269a30e6efbe513e0c17","bottom":"589e579f6994fa9fc99b6e0e0a4da45867fb635005acd0511921191031a8c464"},"consoleErrors":[],"networkRequests":[]},{"viewport":"mobile","passed":true,"localePassed":true,"metrics":{"language":"en","h1Count":1,"h2Count":6,"sectionCount":7,"textLength":3000,"cjkCount":0,"latinLetterCount":2312,"scrollWidth":390,"clientWidth":390,"scrollHeight":6504,"h1Top":148.859375,"bottomInset":375.046875,"brokenImageCount":0,"missingAltCount":0,"fakeControlCount":0,"emptyLinkCount":0},"screenshotHashes":{"top":"56a33d4167e7e71997c5912d2ce2c4413470a95168a58ccd0934c70ecff603e6","middle":"e754e50d75e23d80b76de4fbd9085911ffb0530d601f4fa87cf0572464b6520e","bottom":"ca2d365a9f68e4a378e112dc0b2143b8ec15adaae723b3e7092e69151f3446a0"},"consoleErrors":[],"networkRequests":[]}]},"visualReview":{"schemaVersion":"dodot-automated-design-review-v1","status":"revise","summary":"Faithful to the recall-thread Brief: palette, linear serif+mono surface grammar, a real thread-to-node signature visual, fully evidence-grounded stats, natural bilingual parity, and no forbidden Hero/card-wall/grid-shell patterns. One gating deficiency: the required mobile re-composition is missing — at 390px the hero keeps the same horizontal desktop braid and the CSS only collapses grids to single columns, so the Brief's 90-degree rotation to a vertical top-to-bottom thread ending in a bottom focus node is not implemented.","checks":{"projectFit":{"passed":true,"evidence":"Hero SVG (ct 151-160) is a horizontal braid of session beads receding left-to-right into one lit gold recall node at right-center (node x=920,y=452; beads cx 180..760 increasing opacity; pull-into-node path; mono hero-node-caption) — matches brief heroGeometry/desktopComposition. Palette matches brief exactly (ink #0e1113, gold #f0b75f, teal #7fd4c8, paper #f3efe6, strand #6b7a85). Editorial serif numerals for section numbers and monospaced labels for commands/memory paths as specified. Section sequence equals brief sectionSequence plus hero."},"evidenceBound":{"passed":true,"evidence":"All claims trace to material.json: 95.2% R@5 and 92% token reduction are both branded as project-stated README claims, not independent facts. :3111 server, MCP/hooks/REST, ~/.claude/projects import, 20+ adapters, 17 native skills, and install command npx @agentmemory/agentmemory match extracted facts and readme.commands. A caveat disclaimer accompanies the benchmark section."},"signatureVisual":{"passed":true,"evidence":"The thread-to-node figure really exists in source: hero SVG draws 5 receding two-tone gradient strands and 6 session beads increasing in size/opacity converging into a lit node (r=26 solid fill + r=46 halo + r=52 stroke ring) at right, with a pull-in thread path and a mono hero-node-caption tracking that node back to prior sessions at :3111. Real and specific, not decorative filler."},"mobileTransformation":{"passed":false,"evidence":"Brief requires the horizontal braid to rotate 90 degrees into a vertical stacked thread ending at a bottom focus node at 390px. The only production CSS is @media (max-width:860px) and (max-width:480px), which collapse .hit/.adapt/.bench/.routes/.cmd-row grids to 1fr and reflow .hero-node-caption to static; the hero-figure SVG keeps the identical horizontal viewBox 1200x760 braid with no rotate, no vertical strand/bead stack, and no bottom node. The mandated mobile re-composition preserving the pull-into-node idea is absent."},"localeParity":{"passed":true,"evidence":"zh-CN and en are structurally identical (same 7 data-dodot sections, same 6 numeral kickers, same commands and figures, same stats). Headings and ledes read naturally in each language (zh-CN fully Chinese with latin confined to commands/syntax; en latin only) and commands/memory paths are kept verbatim in both, matching localeContract. No claim, figure, or section diverges."},"antiTemplate":{"passed":true,"evidence":"No generic slogan-over-three-metric-cards (figures shown as labeled project-stated stats, not a 3-card grid), no reusable feature-card wall (routes are thin bordered list grounded in real integration routes), no browser/terminal frame, and no glow/grid/particle background — body background is the flat ink var(--bg) and the V3 radial-gradient backdrops were discarded. Composition is driven by the threaded spine and recall figures."},"noFakeUi":{"passed":true,"evidence":"Every command is real and traceable to readme.commands: npx @agentmemory/agentmemory, npx @agentmemory/agentmemory demo, import-jsonl ~/.claude/projects, upgrade, and npx skills add rohitg00/agentmemory -y. No fake terminal or browser chrome; qa-report records fakeControlCount 0, brokenImageCount 0, emptyLinkCount 0, and no console errors for every locale x viewport."}},"issues":[{"code":"MOBILE_RECOMPOSE_MISSING","severity":"blocking","locale":"all","description":"The Brief's mobileTransformation is not implemented: at 390px the hero remains the same horizontal left-to-right braid built for desktop (same SVG viewBox 1200x760, no rotation), and the media queries only collapse grid columns. The required vertical stacked thread falling to a bottom focus node is absent, so the signature 'pull into node' concept is not re-composed for a phone.","repairInstruction":"Add a phone breakpoint (around 390px) that re-composes the hero vertically: rotate or rebuild the strand/bead sequence to run top-to-bottom ending on a bottom lit recall node, move .hero-node-caption below it, and restack the thread spine, routes, adapters, commands, and figures into a single column so the pull-into-node reads from top to bottom instead of keeping the desktop horizontal braid."}]}},"after":{"qaPassed":true,"designReviewApproved":true,"qaEvidence":{"zh-CN":[{"viewport":"desktop","passed":true,"localePassed":true,"metrics":{"language":"zh-CN","h1Count":1,"h2Count":6,"sectionCount":7,"textLength":1548,"cjkCount":641,"latinLetterCount":536,"scrollWidth":1280,"clientWidth":1280,"scrollHeight":4690,"h1Top":236.3125,"bottomInset":266.921875,"brokenImageCount":0,"missingAltCount":0,"fakeControlCount":0,"emptyLinkCount":0},"screenshotHashes":{"top":"5985320d99c6bc721729d500bc5d0efd4907c6cabb36ecd175b560c51e737b68","middle":"65070ceb39861f4e935eab060d4f9c10255597de42ca029bf5640b2a72419231","bottom":"babb46ddefc2ceadad3c0023980314a721d62667c6a452d0ac139063cc12cf31"},"consoleErrors":[],"networkRequests":[]},{"viewport":"mobile","passed":true,"localePassed":true,"metrics":{"language":"zh-CN","h1Count":1,"h2Count":6,"sectionCount":7,"textLength":1548,"cjkCount":641,"latinLetterCount":536,"scrollWidth":390,"clientWidth":390,"scrollHeight":5843,"h1Top":124.859375,"bottomInset":374.984375,"brokenImageCount":0,"missingAltCount":0,"fakeControlCount":0,"emptyLinkCount":0},"screenshotHashes":{"top":"0a431a0dab4a589caf0c3bdba158a660543e000887363d0930ba632882864314","middle":"0a92c45fe58de84a7dcbc6d63d9651aa19c896f94adf524a9abb91bd4227570b","bottom":"78f75f49617059d7f2acadc0b6043f67e15ead5721c95ea2d8c5c8bbaaa69f9d"},"consoleErrors":[],"networkRequests":[]}],"en":[{"viewport":"desktop","passed":true,"localePassed":true,"metrics":{"language":"en","h1Count":1,"h2Count":6,"sectionCount":7,"textLength":3000,"cjkCount":0,"latinLetterCount":2312,"scrollWidth":1280,"clientWidth":1280,"scrollHeight":4933,"h1Top":113.828125,"bottomInset":266.890625,"brokenImageCount":0,"missingAltCount":0,"fakeControlCount":0,"emptyLinkCount":0},"screenshotHashes":{"top":"b4ec03de74692489646bc519e23c43bad436beba109a2b465e3c9962b792d1d2","middle":"8942ec4fe05ff192f2cf5dcc450d0b60cbdf27fe34dc269a30e6efbe513e0c17","bottom":"589e579f6994fa9fc99b6e0e0a4da45867fb635005acd0511921191031a8c464"},"consoleErrors":[],"networkRequests":[]},{"viewport":"mobile","passed":true,"localePassed":true,"metrics":{"language":"en","h1Count":1,"h2Count":6,"sectionCount":7,"textLength":3000,"cjkCount":0,"latinLetterCount":2312,"scrollWidth":390,"clientWidth":390,"scrollHeight":6465,"h1Top":124.859375,"bottomInset":375.421875,"brokenImageCount":0,"missingAltCount":0,"fakeControlCount":0,"emptyLinkCount":0},"screenshotHashes":{"top":"e156d91b0279d451608a6a78c1b8dd824d27b6d1a1819d534994fa1c0f1d4eb3","middle":"462293ae4bef82bc2bbb53a43826b011116bccaba9f730d0354dadd25865b5cc","bottom":"49842b5ae3f5d6422c7af42f3acc499fd2dc62fea903280985d5a2e1958da0d2"},"consoleErrors":[],"networkRequests":[]}]},"visualReview":{"schemaVersion":"dodot-automated-design-review-v1","status":"approved","summary":"忠实实现 recall-thread Brief：Hero 为水平编织线程拉向单个点亮的召回节点，纯 monospace+serif 表面、精确调色板、全部数字标注为项目自述基准并附声明，双语结构等价且自然，命令/记忆路径逐字保留，无通用 Hero/三卡/卡片墙/设备壳/装饰渐变，且 390px 手机断点使用独立纵向 hero-svg-mobile（自上而下的线程与串珠落入底部召回节点）真正重构了签名视觉，无假控件。七项检查全部通过，无 blocking issue。","checks":{"projectFit":{"passed":true,"evidence":"Hero 桌面 SVG（hero-svg-desktop，viewBox 1200x760）绘制 5 条双色渐变 strands、6 颗随透明度增大的 session beads（cx 180→760）收敛到一个点亮的金色召回节点（x=920,y=452，r=26 实心+r=46 光晕+r=52 圆环），并有 thread pulled into node 路径及 monospace hero-node-caption 把节点追踪回历史会话，完全对应 brief 的 heroGeometry/desktopComposition/signatureVisual。调色板逐字匹配 brief（ink #0e1113、gold #f0b75f、teal #7fd4c8、paper #f3efe6、strand #6b7a85），serif 章节数字 + monospace 命令/记忆路径标签。章节顺序 the-reask-gap/recall-hit/store-and-hooks/reach-adapters/runCommands/recallBenchmark 与 brief sectionSequence 一致。选中的 selectedDirectionId=recall-thread，无 index-fanout/ledger-replay 方向泄漏。"},"evidenceBound":{"passed":true,"evidence":"全部事实均可在 material.json 中找到依据：95.2% R@5 与 92% token 减少均在页面中标为『项目自述基准/README claim, not an independent test』并附 caveat 声明（非独立事实）；:3111 内存服务器、MCP/hooks/REST、~/.claude/projects 导入、20+ 适配器、17 原生技能、安装命令 npx @agentmemory/agentmemory 均对应提取事实与 readme.commands；0.9.29/Cursor 插件、Devin/DeepSeek harness 等未虚构。reach-adapters 标签（Claude Code/Cursor/Codex CLI/Gemini CLI/Hermes/Devin/OpenClaw）皆有 README/lifecycle 依据。"},"signatureVisual":{"passed":true,"evidence":"thread-to-node 图在源码中真实存在且专属：桌面 SVG 用 5 条渐变 strands + 6 颗递增串珠 + 高亮召回节点 + pull-in 路径组成『编织线程拉向单个点亮节点』；底部 recall-hit 与 recallBenchmark 的 SVG 图继续沿用 hit-node 与 re-explaining-falls-away 曲线，图中含 :3111 命中节点标注。该图形是项目机制的真实表达，并非装饰性渐变/粒子填充。"},"mobileTransformation":{"passed":true,"evidence":"Brief 要求手机端把水平编织旋转 90° 成纵向、自上而下落入底部焦点节点。源码实现了独立 hero-svg-mobile（viewBox 390x1080），在 @media(max-width:480px) 下 hero-svg-desktop display:none 而 hero-svg-mobile display:block：5 条 strands 从顶部向下流动、6 颗串珠 cy 140→720 递增，终点为底部点亮召回节点（cx=198,cy=880，r=30 实心+r=52 光晕+r=60 圆环）并有 thread 向下拉入底节点的路径。@media(max-width:860px) 使 .hit/.adapt/.bench/.routes/.cmd-row 塌缩为单列，480px 下 .hero-node-caption 移到底部，保留了『拉入节点』的意图，满足 390px 重构要求。"},"localeParity":{"passed":true,"evidence":"zh-CN 与 en 页面结构完全一致：相同的 7 个 data-dodot-section、相同的 6 个数字 kicker、相同命令与 SVG 图、相同统计数字。两种语言文案自然（zh-CN 全中文，latin 仅用于命令/路径/语法；en 全英文），命令与记忆路径在双语中逐字保留，符合 localeContract。zh-CN 无中文码点混入命令，en 无英文出现于 label 外的中文。无任一章节/事实/图偏移。"},"antiTemplate":{"passed":true,"evidence":"未回退到通用 Hero/三卡/卡片墙/设备壳/装饰网格：统计以带来源标注的项目自述数字呈现而非 slogan-over-3-metric-cards；store-and-hooks 的三个 ROUTE 是扎根于真实集成路径（MCP/hooks/REST）的细线列表而非可复用浮空卡片墙；无浏览器/终端外壳框（V3 的设备外壳与 radial-gradient 背景已被丢弃，body 用纯色 --bg）；无发光/网格/粒子作为装饰填充。排版与组成由线程脊柱与召回图驱动。"},"noFakeUi":{"passed":true,"evidence":"每个命令均可追溯 readme.commands：npx @agentmemory/agentmemory、demo、import-jsonl ~/.claude/projects、upgrade、以及 npx skills add rohitg00/agentmemory -y 全部真实存在。无假终端或浏览器 chrome。qa-report.json 对每个 locale x viewport 均记录 fakeControlCount 0、brokenImageCount 0、missingAltCount 0、emptyLinkCount 0、无 consoleErrors、无 networkRequests。"}},"issues":[]}},"changedFiles":["artifact/zh-CN.html","artifact/en.html"]}],"provenance":{"executor":"opencode","model":"deepseek/deepseek-v4-flash","opencodeVersion":"1.1.58","promptVersion":"dodot-showcase-deepseek-v4","inputHash":"a8ea339010a37802c7033e5900158aba1908ff2c9782ecfa81b41381aa4fdcd5","artifactHash":"fe6a9c2829cd3a6aea023b87b7a5da939d2eeaee459a99ac5c162b35cba1b45d","designSkill":"dodot-showcase-design","historicalSeed":{"source":"backup_database","sourceCanonicalKey":"github:rohitg00/agentmemory","assetId":"cmsv8evy5006jbnqtd21i8eko","sourceSchemaVersion":"localized-static-html-v3","sourceGeneratorVersion":"opencode-1.1.58:deepseek/deepseek-v4-flash","sourceArtifactHash":"e7f9a6d7bbf0ed5e837ccc8c93b0f77cd4d7a278c05cb4d06c6013d2ba513606","disposition":"validate_and_upgrade","seedPayloadHash":"c20c82a8dc404c5dbd740ee0658c489edb928725b0e7d51723000948dfffd8c4"}},"qaEvidence":{"zh-CN":[{"viewport":"desktop","passed":true,"localePassed":true,"metrics":{"language":"zh-CN","h1Count":1,"h2Count":6,"sectionCount":7,"textLength":1548,"cjkCount":641,"latinLetterCount":536,"scrollWidth":1280,"clientWidth":1280,"scrollHeight":4690,"h1Top":236.3125,"bottomInset":266.921875,"brokenImageCount":0,"missingAltCount":0,"fakeControlCount":0,"emptyLinkCount":0},"screenshotHashes":{"top":"5985320d99c6bc721729d500bc5d0efd4907c6cabb36ecd175b560c51e737b68","middle":"65070ceb39861f4e935eab060d4f9c10255597de42ca029bf5640b2a72419231","bottom":"babb46ddefc2ceadad3c0023980314a721d62667c6a452d0ac139063cc12cf31"},"consoleErrors":[],"networkRequests":[]},{"viewport":"mobile","passed":true,"localePassed":true,"metrics":{"language":"zh-CN","h1Count":1,"h2Count":6,"sectionCount":7,"textLength":1548,"cjkCount":641,"latinLetterCount":536,"scrollWidth":390,"clientWidth":390,"scrollHeight":5843,"h1Top":124.859375,"bottomInset":374.984375,"brokenImageCount":0,"missingAltCount":0,"fakeControlCount":0,"emptyLinkCount":0},"screenshotHashes":{"top":"0a431a0dab4a589caf0c3bdba158a660543e000887363d0930ba632882864314","middle":"0a92c45fe58de84a7dcbc6d63d9651aa19c896f94adf524a9abb91bd4227570b","bottom":"78f75f49617059d7f2acadc0b6043f67e15ead5721c95ea2d8c5c8bbaaa69f9d"},"consoleErrors":[],"networkRequests":[]}],"en":[{"viewport":"desktop","passed":true,"localePassed":true,"metrics":{"language":"en","h1Count":1,"h2Count":6,"sectionCount":7,"textLength":3000,"cjkCount":0,"latinLetterCount":2312,"scrollWidth":1280,"clientWidth":1280,"scrollHeight":4933,"h1Top":113.828125,"bottomInset":266.890625,"brokenImageCount":0,"missingAltCount":0,"fakeControlCount":0,"emptyLinkCount":0},"screenshotHashes":{"top":"b4ec03de74692489646bc519e23c43bad436beba109a2b465e3c9962b792d1d2","middle":"8942ec4fe05ff192f2cf5dcc450d0b60cbdf27fe34dc269a30e6efbe513e0c17","bottom":"589e579f6994fa9fc99b6e0e0a4da45867fb635005acd0511921191031a8c464"},"consoleErrors":[],"networkRequests":[]},{"viewport":"mobile","passed":true,"localePassed":true,"metrics":{"language":"en","h1Count":1,"h2Count":6,"sectionCount":7,"textLength":3000,"cjkCount":0,"latinLetterCount":2312,"scrollWidth":390,"clientWidth":390,"scrollHeight":6465,"h1Top":124.859375,"bottomInset":375.421875,"brokenImageCount":0,"missingAltCount":0,"fakeControlCount":0,"emptyLinkCount":0},"screenshotHashes":{"top":"e156d91b0279d451608a6a78c1b8dd824d27b6d1a1819d534994fa1c0f1d4eb3","middle":"462293ae4bef82bc2bbb53a43826b011116bccaba9f730d0354dadd25865b5cc","bottom":"49842b5ae3f5d6422c7af42f3acc499fd2dc62fea903280985d5a2e1958da0d2"},"consoleErrors":[],"networkRequests":[]}]}},"generatorVersion":"opencode-1.1.58:deepseek/deepseek-v4-flash:dodot-showcase-deepseek-v4","publishedAt":"2026-08-20T14:39:01.904Z"},"rawExtraction":{"projectRevisionId":"cmt2zh2o801gl6271kgb95x2i","projectVersion":1,"sources":[{"canonicalKey":"github:rohitg00/agentmemory","role":"primary","revisionVersion":1,"url":"https://github.com/rohitg00/agentmemory"}],"facts":[{"key":"repository_owner","value":"rohitg00","evidence":[{"path":"repository.owner","excerpt":"rohitg00"}]},{"key":"repository_name","value":"agentmemory","evidence":[{"path":"repository.name","excerpt":"agentmemory"}]},{"key":"description","value":"#1 Persistent memory for AI coding agents based on real-world benchmarks","evidence":[{"path":"repository.description","excerpt":"#1 Persistent memory for AI coding agents based on real-world benchmarks"}]},{"key":"primary_language","value":"TypeScript","evidence":[{"path":"repository.language","excerpt":"TypeScript"}]},{"key":"license","value":"Apache-2.0","evidence":[{"path":"repository.license","excerpt":"Apache-2.0"}]},{"key":"topics","value":"agentmemory, agents, ai, claude, claudecode, codex, copilot, cursor, genai, harness, hermes, memory, openclaw","evidence":[{"path":"repository.topics","excerpt":"agentmemory, agents, ai, claude, claudecode, codex, copilot, cursor, genai, harness, hermes, memory, openclaw"}]},{"key":"source_count","value":"1","evidence":[{"path":"projectRevision.sources","excerpt":"github:rohitg00/agentmemory"}]}],"taxonomy":{"category":"development_tools","subcategory":"data_ai_tools","confidence":0.65,"evidence":[{"path":"repository.description","excerpt":"#1 Persistent memory for AI coding agents based on real-world benchmarks"}]},"lifecycle":{"stage":"prototype","confidence":0.65,"evidence":[{"path":"readmeExcerpt","excerpt":"

$\"agentmemory:$

Your coding agent remembers everything. No more re-explaining. Built on iii engine
Persistent memory for Claude Code, GitHub Copilot CLI, Cursor, Gemini CLI, Codex CLI, Hermes, OpenClaw, pi, OpenCode, and any MCP client.

English |