[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"article-mempalace-100-percent-claim-scrutiny-zh":3,"tags-mempalace-100-percent-claim-scrutiny-zh":33,"related-lang-mempalace-100-percent-claim-scrutiny-zh":49,"related-posts-mempalace-100-percent-claim-scrutiny-zh":53,"series-ai-agent-4f005451-c02c-4f08-90c5-c1f94b5c374a":90},{"id":4,"title":5,"content":6,"summary":7,"source":8,"source_url":9,"author":10,"image_url":11,"keywords":12,"language":21,"translated_content":10,"views":22,"is_premium":23,"created_at":24,"updated_at":24,"cover_image":11,"published_at":25,"rewrite_status":26,"rewrite_error":10,"rewritten_from_id":27,"slug":28,"category":29,"related_article_id":30,"status":31,"google_indexed_at":32,"x_posted_at":10,"tweet_text":10,"title_rewritten_at":10,"title_original":10,"key_takeaways":10,"topic_cluster_id":10,"embedding":10,"is_canonical_seed":23},"4f005451-c02c-4f08-90c5-c1f94b5c374a","MemPalace 的 100% 記憶宣稱被拆穿","\u003Cp>\u003Ca href=\"https:\u002F\u002Fgithub.com\u002Fmilla-jovovich\u002Fmempalace\" target=\"_blank\" rel=\"noopener\">MemPalace\u003C\u002Fa> 48 小時內衝到 1.1 萬 GitHub stars。這數字很猛，代表大家真的在找 AI 記憶工具。問題是，它主打的 \u003Ca href=\"https:\u002F\u002Fgithub.com\u002Fucsd-ml\u002FLongMemEval\" target=\"_blank\" rel=\"noopener\">LongMemEval\u003C\u002Fa> 100% 分數，後來在啟用壓縮後掉到 84.2%。\u003C\u002Fp>\u003Cp>講白了，專案本身不假。誇張的是宣傳。這種差距很重要，因為 AI 工具圈最常見的坑，就是 demo 很漂亮，實戰很普通。\u003C\u002Fp>\u003Cp>我覺得這篇最值得看的，不是誰被打臉。是你能不能看懂，哪些數字是工程結果，哪些只是行銷話術。\u003C\u002Fp>\u003Ch2>MemPalace 到底在做什麼\u003C\u002Fh2>\u003Cp>MemPalace 的核心想法不複雜。它不是把聊天紀錄平鋪在一個長 log 裡。它把記憶切成 wing、hall、room。這套設計借用「記憶宮殿」概念，讓系統用空間結構去整理資料。\u003C\u002Fp>\n\u003Cfigure class=\"my-6\">\u003Cimg src=\"https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1775579095468-akte.png\" alt=\"MemPalace 的 100% 記憶宣稱被拆穿\" class=\"rounded-xl w-full\" loading=\"lazy\" \u002F>\u003C\u002Ffigure>\n\u003Cp>它底層用 \u003Ca href=\"https:\u002F\u002Fwww.trychroma.com\u002F\" target=\"_blank\" rel=\"noopener\">ChromaDB\u003C\u002Fa> 做檢索。設定檔和中繼資料則交給 \u003Ca href=\"https:\u002F\u002Fgithub.com\u002Fyaml\u002Fpyyaml\" target=\"_blank\" rel=\"noopener\">PyYAML\u003C\u002Fa>。另外，它還有 \u003Ca href=\"https:\u002F\u002Fmodelcontextprotocol.io\u002F\" target=\"_blank\" rel=\"noopener\">Model Context Protocol\u003C\u002Fa> 的 MCP s\u003Ca href=\"\u002Fnews\u002Fserver-learning-hardened-federated-learning-zh\">erve\u003C\u002Fa>r。這代表它能接到 Claude、Cursor 這類工具。\u003C\u002Fp>\u003Cp>這種 local-first 設計，對台灣開發者很有感。很多 AI 記憶產品都先把資料丟雲端。方便是方便，但資料控制權也一起送出去。MemPalace 至少把記憶留在本機，這點很實際。\u003C\u002Fp>\u003Cul>\u003Cli>48 小時 GitHub stars：11,000+\u003C\u002Fli>\u003Cli>原始宣稱：LongMemEval 100%\u003C\u002Fli>\u003Cli>壓縮後實測：84.2%\u003C\u002Fli>\u003Cli>R@5 檢索成績：96.6%\u003C\u002Fli>\u003Cli>MCP 工具數：19 個\u003C\u002Fli>\u003C\u002Ful>\u003Ch2>100% 分數為何站不住腳\u003C\u002Fh2>\u003Cp>LongMemEval 不是假 benchmark。它來自 UC San Diego，總共 500 題，測的是長期記憶能力。題目涵蓋多種記憶任務，不是隨便跑個 embedding 相似度就能混過去。\u003C\u002Fp>\u003Cp>但問題出在呈現方式。外部檢查發現，最後三題錯誤是先手動修掉，再重跑同一份資料集。這就很像先看答案再考一次。分數會漂亮，但不能證明系統真的泛化。\u003C\u002Fp>\u003Cblockquote>“The first principle is that you must not fool yourself and you are the easiest person to fool.” — Richard Feynman\u003C\u002Fblockquote>\u003Cp>費曼這句話很適合這次事件。\u003Ca href=\"\u002Fnews\u002Fhierarchical-planning-latent-world-models-zh\">模型\u003C\u002Fa>可以是真的，repo 可以是真的，分數還是可能很會騙人。AI 圈最怕的，不是沒成果，是把 demo 當結論。\u003C\u002Fp>\u003Cp>還有一個更細的問題。96.6% 的 R@5，看起來像是 ChromaDB 預設檢索撐出來的。那比較像最近鄰搜尋，不等於整套 memory palace 架構真的做到 lossless。等到 AAAK 壓縮真的上場，成績就掉到 84.2%。\u003C\u002Fp>\u003Ch2>跟其他記憶系統比，差在哪\u003C\u002Fh2>\u003Cp>如果只看 LongMemEval，MemPalace 其實還算能打。只是它沒有宣稱的那麼神。把數字攤開來看，它比較像一個不錯的本機記憶原型，而不是什麼滿分怪物。\u003C\u002Fp>\n\u003Cfigure class=\"my-6\">\u003Cimg src=\"https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1775579101284-zesr.png\" alt=\"MemPalace 的 100% 記憶宣稱被拆穿\" class=\"rounded-xl w-full\" loading=\"lazy\" \u002F>\u003C\u002Ffigure>\n\u003Cp>先看幾個對照組。這些數字來自公開 repo 或評測說明。你會發現，MemPalace 的原始檢索不差，但壓縮後就不太穩。\u003C\u002Fp>\u003Cul>\u003Cli>\u003Ca href=\"https:\u002F\u002Fgithub.com\u002Fmastra-ai\u002Fmastra\" target=\"_blank\" rel=\"noopener\">Mastra\u003C\u002Fa>：94.87%\u003C\u002Fli>\u003Cli>\u003Ca href=\"https:\u002F\u002Fgithub.com\u002Fomega-memory\u002Fomega\" target=\"_blank\" rel=\"noopener\">OMEGA\u003C\u002Fa>：95.4%\u003C\u002Fli>\u003Cli>\u003Ca href=\"https:\u002F\u002Fgithub.com\u002Fagentmemory\u002Fagentmemory\" target=\"_blank\" rel=\"noopener\">agentmemory\u003C\u002Fa>：96.2%\u003C\u002Fli>\u003Cli>MemPalace 原始檢索：96.6% R@5\u003C\u002Fli>\u003Cli>MemPalace 壓縮後：84.2%\u003C\u002Fli>\u003C\u002Ful>\u003Cp>這組比較很直白。MemPalace 的檢索能力有料，但 100% 這種說法太滿。你如果是做 agent、RAG、或個人知識庫，真正該看的不是滿分，而是壓縮後還剩多少。\u003C\u002Fp>\u003Cp>架構上也有差異。很多記憶系統，本質上就是向量資料庫加幾條規則。MemPalace 想模仿人類回憶路徑，用空間結構整理資訊。這想法比較有趣，但也更容易被資料分布打臉。\u003C\u002Fp>\u003Ch2>誰真的做了這個專案\u003C\u002Fh2>\u003Cp>Milla Jovovich 的參與不是假的。她的 GitHub bio 寫自己是 MemPalace 的 architect。她的公開社群足跡也對得上。這表示她不是掛名而已，至少有實際參與方向設計。\u003C\u002Fp>\u003Cp>Ben Sigman 的貼文也透露了更多細節。他提到是用 Claude 做出來的，還開玩笑講了 “Multipass”。這很明顯是在說 \u003Ca href=\"https:\u002F\u002Fwww.anthropic.com\u002Fclaude-code\" target=\"_blank\" rel=\"noopener\">Claude Code\u003C\u002Fa> 幫忙寫了不少程式。這沒什麼好羞恥的，現在很多團隊都這樣做。\u003C\u002Fp>\u003Cp>這件事真正有趣的地方，在於它改寫了誰能做軟體。以前你可能要先會寫完整系統，才有資格做產品。現在只要 workflow 對，\u003Ca href=\"\u002Fnews\u002Fbas-llm-confidence-abstain-decisions-zh\">LLM\u003C\u002Fa> 夠穩，一個沒有傳統工程背景的人也能把工具做出來。\u003C\u002Fp>\u003Cp>但這也提醒我們一件事。AI 輔助開發很強，不代表評測就可以亂講。專案可以是真的，數字也可以是真的，兩者還是可能一起出現問題。\u003C\u002Fp>\u003Ch2>這件事對開發者有什麼用\u003C\u002Fh2>\u003Cp>如果你在做 agent 或記憶系統，MemPalace 值得看的是設計，不是那個 100%。它的空間化記憶模型，對某些工作流可能比平面 log 更好用。尤其是要人工檢查、刪除、或重組記憶時。\u003C\u002Fp>\u003Cp>但你也要更小心評測。只要幾題能手動補，分數就能往上拉。那種測法很容易把工程品質包裝成成果。這條原則不只適用 open source，也適用新創 demo 和內部 prototype。\u003C\u002Fp>\u003Cp>我自己的判斷很簡單。MemPalace 是一個有料的原型。只是它被包在太滿的宣稱裡。真正值得學的，是 local-first、MCP、空間記憶這三件事，不是那個 100%。\u003C\u002Fp>\u003Cp>接下來最該問的，不是它有沒有上過 1.1 萬 stars。是它換新模型、換新資料、換新任務後，還能不能維持同樣水準。這才是開發者真正要看的地方。\u003C\u002Fp>\u003Ch2>AI 記憶工具接下來會怎麼走\u003C\u002Fh2>\u003Cp>AI 記憶這條線，現在很像早期 RAG。大家都在找更好的檢索、更好的壓縮、更好的上下文管理。問題是，很多產品都先講故事，再補技術細節。\u003C\u002Fp>\u003Cp>台灣團隊如果要做這類工具，我會建議先把三件事做扎實。第一，資料留在哪裡。第二，壓縮後還剩多少可用資訊。第三，評測是不是能被輕易修答案。這三件事比宣傳頁面重要很多。\u003C\u002Fp>\u003Cp>MemPalace 給的訊號很清楚。它證明 local memory 有市場，也證明一個好看的分數，不能直接等於好產品。下次你看到 100% 這種字眼，先問一句：壓縮開了嗎？資料換了嗎？新樣本跑過嗎？\u003C\u002Fp>\u003Cp>我猜接下來 6 到 12 個月，這類工具會更重視可驗證性。誰能把記憶結構、檢索流程、和評測方法講清楚，誰就比較容易留下來。說真的，這比喊滿分有用多了。\u003C\u002Fp>","MemPalace 48 小時衝破 1.1 萬 GitHub stars，但 LongMemEval 的 100% 記憶分數在啟用壓縮後掉到 84.2%。專案本身有料，宣傳數字卻太滿。","oracore-original","https:\u002F\u002Fgithub.com\u002Fmilla-jovovich\u002Fmempalace",null,"https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1775579095468-akte.png",[13,14,15,16,17,18,19,20],"MemPalace","AI 記憶","LongMemEval","MCP","ChromaDB","Claude Code","LLM","本機優先","zh",4,false,"2026-04-07T16:24:36.390934+00:00","2026-04-07T16:24:36.219+00:00","done","58d0006b-cff1-4fa5-86f0-47b1d08a7741","mempalace-100-percent-claim-scrutiny-zh","ai-agent","406aeb11-f9c0-428a-9160-e983f7977d2e","published","2026-04-08T09:00:48.019+00:00",[34,36,37,39,41,43,45,47],{"name":14,"slug":35},"ai-記憶",{"name":20,"slug":20},{"name":16,"slug":38},"mcp",{"name":15,"slug":40},"longmemeval",{"name":18,"slug":42},"claude-code",{"name":13,"slug":44},"mempalace",{"name":19,"slug":46},"llm",{"name":17,"slug":48},"chromadb",{"id":30,"slug":50,"title":51,"language":52},"mempalace-100-percent-claim-scrutiny-en","MemPalace’s 100% memory claim gets checked","en",[54,60,66,72,78,84],{"id":55,"slug":56,"title":57,"cover_image":58,"image_url":58,"created_at":59,"category":29},"e7874ed9-592f-4e06-b7b7-ab733fe779db","claude-agent-dreaming-outcomes-multiagent-zh","Claude 幫 Agent 加了做夢功能","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1778868642412-7woy.png","2026-05-15T18:10:24.427608+00:00",{"id":61,"slug":62,"title":63,"cover_image":64,"image_url":64,"created_at":65,"category":29},"38406a12-f833-4c69-ae22-99c31f03dd52","switch-ai-outputs-markdown-to-html-zh","怎麼把 AI 輸出改成 HTML","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1778743243861-8901.png","2026-05-14T07:20:21.545364+00:00",{"id":67,"slug":68,"title":69,"cover_image":70,"image_url":70,"created_at":71,"category":29},"c7c69fe4-97e3-4edf-a9d6-a79d0c4495b4","anthropic-cat-wu-proactive-ai-assistants-zh","Cat Wu 談 Claude 的主動式 AI","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1778735455993-gnw7.png","2026-05-14T05:10:30.453046+00:00",{"id":73,"slug":74,"title":75,"cover_image":76,"image_url":76,"created_at":77,"category":29},"e1d6acda-fa49-4514-aa75-709504be9f93","how-to-run-hermes-agent-on-discord-zh","如何在 Discord 執行 Hermes Agent","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1778724655796-cjul.png","2026-05-14T02:10:34.362605+00:00",{"id":79,"slug":80,"title":81,"cover_image":82,"image_url":82,"created_at":83,"category":29},"4104fa5f-d95f-45c5-9032-99416cf0365c","why-ragflow-is-the-right-open-source-rag-engine-to-self-host-zh","為什麼 RAGFlow 是最適合自架的開源 RAG 引擎","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1778674262278-1630.png","2026-05-13T12:10:23.762632+00:00",{"id":85,"slug":86,"title":87,"cover_image":88,"image_url":88,"created_at":89,"category":29},"7095f05c-34f5-469f-a044-2525d2010ce9","how-to-add-temporal-rag-in-production-zh","如何在正式環境加入 Temporal RAG","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1778667053844-osvs.png","2026-05-13T10:10:30.930982+00:00",[91,96,101,106,111,116,121,126,131,136],{"id":92,"slug":93,"title":94,"created_at":95},"4ae1e197-1d3d-4233-8733-eafe9cb6438b","claude-now-uses-your-pc-to-finish-tasks-zh","Claude 開始幫你操作電腦","2026-03-26T07:20:48.457387+00:00",{"id":97,"slug":98,"title":99,"created_at":100},"5bede67f-e21c-413d-9ab8-54a3c3d26227","googles-2026-ai-agent-report-decoded-zh","Google 2026 AI Agent 報告解讀","2026-03-26T11:15:22.651956+00:00",{"id":102,"slug":103,"title":104,"created_at":105},"2987d097-563f-46c7-b76f-b558d8ef7c2b","kimi-k25-review-stronger-still-not-legend-zh","Kimi K2.5 評測：更強，但還不是神作","2026-03-27T07:15:55.277513+00:00",{"id":107,"slug":108,"title":109,"created_at":110},"95c9053b-e3f4-4cb5-aace-5c54f4c9e044","claude-code-controls-mac-desktop-zh","Claude Code 也能操控 Mac 了","2026-03-28T03:01:58.58121+00:00",{"id":112,"slug":113,"title":114,"created_at":115},"dc58e153-e3a8-4c06-9b96-1aa64eabbf5f","cloudflare-100x-faster-ai-agent-sandbox-zh","Cloudflare 的 AI 沙箱跑超快","2026-03-28T03:09:44.142236+00:00",{"id":117,"slug":118,"title":119,"created_at":120},"1c8afc56-253f-47a2-979f-1065ff072f2a","openai-backs-isara-agent-swarm-bet-zh","OpenAI 挺 Isara 的 agent swarm …","2026-03-28T03:15:27.513155+00:00",{"id":122,"slug":123,"title":124,"created_at":125},"7379b422-576e-45df-ad5a-d57a0d9dd467","openai-plan-automated-ai-researcher-zh","OpenAI 想做自動化 AI 研究員","2026-03-28T03:17:42.090548+00:00",{"id":127,"slug":128,"title":129,"created_at":130},"48c9889e-86df-450b-a356-e4a4b7c83c5b","harness-engineering-ai-agent-reliability-2026-zh","駕馭工程：從「馬具」到「作業系統」，AI Agent 可靠性的終極密碼","2026-03-31T06:42:53.556721+00:00",{"id":132,"slug":133,"title":134,"created_at":135},"e41546b8-ba9e-455f-9159-88d4614ad711","openai-codex-plugin-claude-code-zh","OpenAI 把 Codex 放進 Claude Code","2026-04-01T09:21:54.687617+00:00",{"id":137,"slug":138,"title":139,"created_at":140},"96d8e8c8-1edd-475d-9145-b1e7a1b02b65","mcp-explained-from-prompts-to-production-zh","MCP 怎麼把提示詞變工作流","2026-04-01T09:24:39.321274+00:00"]