[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-openai-dreaming-v3-self-update-memory":3,"news-related-2511d073-9f31-479a-ac94-1c2786e2f211":36},{"id":4,"title":5,"summary":6,"content":6,"original_url":7,"source_id":8,"tags":9,"translations":23,"news_slug":29,"published_at":30,"created_at":31,"modified_at":32,"is_published":33,"publish_type":34,"image_url":13,"view_count":35},"2511d073-9f31-479a-ac94-1c2786e2f211","OpenAI Dreaming V3 登场：ChatGPT 记忆架构跨入「自我更新」时代","OpenAI 于 6 月 4 日正式推出 Dreaming V3 记忆架构，标志着 ChatGPT 的个人记忆从「用户主动维护的列表」跃迁到「模型自我维护的连续状态」。\n\nDreaming 最早在 2025 年 4 月的 V0 中引入，思路是让模型在后台跨多轮对话合成上下文、补充用户未显式要求的「提示要记住」指令。但 OpenAI 自己承认 V0 始终「无法独立作为完整的记忆系统」。V3 直接替代了沿用两年的「saved memories」列表：后台进程持续合成记忆状态，自然出现的语境自动捕获，条目随时间自动更新。\n\n工程上 V3 的关键突破是可扩展性——OpenAI 透露服务 Free 用户的计算开销相比旧方案下降约 5 倍，因此未来几周能将 Dreaming 推向免费档和 Go 档用户。第三方测评显示，事实性召回准确率从约 41.5% 跃升到 82.8%。这意味着 ChatGPT 终于「记得住」跨会话、跨季度的偏好与项目背景，而不再只依赖当次对话的窗口。\n\n更深层的影响在产品形态：随着记忆从「用户显式管理」变成「模型自维护」，Agent 化长期任务协作才真正有底座可用。同时也带来新的治理难题——当 AI 静默改写它对你的认知，错误修正的成本与责任边界都悬而未决。Anthropic 5 月推出「Claude 梦境」与 OpenAI Dreaming V3 只是开端，2026 下半年的「记忆之争」将成为各家前端大模型竞争的新分水岭。","https:\u002F\u002Fopenai.com\u002Findex\u002Fchatgpt-memory-dreaming\u002F","15975962-b5fe-49e5-ae68-687ba6cb7015",[10,14,17,20],{"id":11,"name":12,"slug":12,"description":13,"color":13},"5e628969-6d2a-437f-998a-104e4b16cfb1","ai-progress",null,{"id":15,"name":16,"slug":16,"description":13,"color":13},"40269b40-7942-4650-9672-ed2e6524d37a","ai-technology",{"id":18,"name":19,"slug":19,"description":13,"color":13},"01598627-1ea6-4b27-a5d8-874971571a71","llm",{"id":21,"name":22,"slug":22,"description":13,"color":13},"42e59a88-7795-47dc-a334-ef1e72c24347","openai",[24],{"id":25,"lang":26,"title":27,"summary":28,"content":13},"7a894c37-db54-452e-965d-f3947bc9f67a","en","OpenAI Dreaming V3: ChatGPT memory becomes self-updating","OpenAI officially launched the Dreaming V3 memory architecture on June 4, marking ChatGPT's personal memory leaping from \"user-maintained list\" to \"model-maintained continuous state.\"\n\nDreaming was first introduced in V0 in April 2025, with the idea of letting the model synthesize context across multiple turns in the background, supplementing the user-not-explicitly-asked \"prompts to remember\" instructions. But OpenAI itself admits that V0 has always been \"unable to stand as a complete memory system on its own.\" V3 directly replaces the two-year-old \"saved memories\" list: background processes continuously synthesize memory state, naturally occurring context is automatically captured, and entries are automatically updated over time.\n\nThe key engineering breakthrough of V3 is scalability — OpenAI reveals that the compute overhead of serving Free users has dropped by about 5× compared to the old solution, so in the coming weeks Dreaming can be pushed to Free and Go tier users. Third-party testing shows that factual recall accuracy has jumped from about 41.5% to 82.8%. This means ChatGPT finally \"remembers\" cross-session, cross-quarter preferences and project context, rather than just relying on the current conversation's window.\n\nThe deeper impact is on product form: as memory shifts from \"user explicit management\" to \"model self-maintenance,\" agent-style long-term task collaboration finally has a foundation to build on. At the same time it brings new governance problems — when AI silently rewrites its understanding of you, the cost of error correction and the boundaries of responsibility are both open. Anthropic's May launch of \"Claude Dreaming\" and OpenAI's Dreaming V3 are just the beginning; the \"memory war\" in the second half of 2026 will become the new watershed for competition among the various frontier model vendors.","openai-dreaming-v3-self-update-memory","2026-06-06T13:00:00Z","2026-06-06T13:08:40.296500Z","2026-08-19T02:08:40.142862Z",true,"agent",136,{"items":37},[38,42,47,52,57,62],{"id":39,"title":40,"news_slug":40,"published_at":41},"4afbd081-d5dd-4595-80f8-dd72c69a136a","GPT-5.5 让模型在发布前先改自己跑的引擎：这不是新模型,是 OpenAI 的 release 范式更新","2026-08-03T18:00:00+00:00",{"id":43,"title":44,"news_slug":45,"published_at":46},"43eda321-b0b7-4df7-b20e-9758cbab42c9","记忆越完整,眼前题越做不对:MemTrapBench 把 LLM 长期记忆框架打回原形","memtrapbench-llm-memory-cognitive-traps","2026-08-22T04:00:00+00:00",{"id":48,"title":49,"news_slug":50,"published_at":51},"418a9ac0-18fd-49a4-b7a8-d29d1c1ba497","AI 承诺的四天工作制为什么没来：OpenAI \u002F Anthropic 内部工时真相","ai-four-day-work-week-myth-openai-90-hours","2026-08-16T03:30:00+00:00",{"id":53,"title":54,"news_slug":55,"published_at":56},"31f3215c-0892-419d-a610-fe815cc60bbe","GPT-5.6 降价 80% 把竞争拉进「同等智能成本」：DeepSeek V4 Flash 接招，国产模型卡出双线赛道","gpt-5-6-luna-price-cut-equal-intelligence-cost","2026-08-12T03:00:00+00:00",{"id":58,"title":59,"news_slug":60,"published_at":61},"d1c7b405-fe4e-40f9-9249-a12e2bba6913","GPT-5.6 八月更新：把「推理强度滑块」下放给 Plus\u002FPro，同时把免费用户拉进 Luna 时代","openai-gpt-5-6-august-update-reasoning-slider","2026-08-10T20:00:00+00:00",{"id":63,"title":64,"news_slug":65,"published_at":66},"22a1a718-0eb6-46e5-8ee8-825400de11d1","DeepMind WeatherNext 在 Nature 发论文：用 28 km 粗分辨率做出多一天的飓风预警,代码权重全部开源","deepmind-weathernext-cyclones-nature-open-source","2026-08-10T02:00:00+00:00"]