From 059febaba77f2db50ec3ee49ba1a2240b291a557 Mon Sep 17 00:00:00 2001 From: yi-ge Date: Tue, 28 Jul 2026 10:30:10 +0800 Subject: [PATCH 1/5] docs(concepts): P0 optimization - fix broken links and add EN auto-dream --- docs/04-concepts/10-task-orchestration.md | 2 +- docs/04-concepts/11-auto-dream.md | 6 +++--- 2 files changed, 4 insertions(+), 4 deletions(-) diff --git a/docs/04-concepts/10-task-orchestration.md b/docs/04-concepts/10-task-orchestration.md index a5f8977..fb61dc7 100644 --- a/docs/04-concepts/10-task-orchestration.md +++ b/docs/04-concepts/10-task-orchestration.md @@ -120,5 +120,5 @@ DesireCore 的**智能任务编排引擎**就是你的"项目经理"。当你提 ## 下一步 - 想了解步骤的可控性？请阅读 [固化/灵活/人闸门](./step-types) -- 想了解多智能体如何协作？请阅读用户指南中的 [跨智能体协作](../user-guide/delegation/cross-agent) +- 想了解多智能体如何协作？请阅读用户指南中的 [跨智能体协作](../02-user-guide/04-delegation/06-cross-agent) - 想了解任务的回执系统？请阅读 [回执系统](./receipt-system) diff --git a/docs/04-concepts/11-auto-dream.md b/docs/04-concepts/11-auto-dream.md index db4c99c..6c8c1fe 100644 --- a/docs/04-concepts/11-auto-dream.md +++ b/docs/04-concepts/11-auto-dream.md @@ -58,7 +58,7 @@ Auto Dream 的工作分为四个阶段： ### 记忆生命周期（含 Dream 阶段） -在 [自动学习与遗忘](../user-guide/memory/auto-learning) 中，我们介绍了记忆的五阶段生命周期。加入 Auto Dream 后，完整的生命周期变为： +在 [自动学习与遗忘](../02-user-guide/07-memory/04-auto-learning) 中，我们介绍了记忆的五阶段生命周期。加入 Auto Dream 后，完整的生命周期变为： ``` 活跃 → 近期 → 归档 → 💤 Dream → 压缩 → 清理 @@ -120,6 +120,6 @@ Auto Dream 不会删除任何原始记忆。即使你对整合结果不满意， ## 下一步 -- 想了解记忆的完整工作原理？请阅读用户指南中的 [智能体如何记住你](../user-guide/memory/how-memory-works) -- 想了解普通遗忘机制？请阅读 [自动学习与遗忘](../user-guide/memory/auto-learning) +- 想了解记忆的完整工作原理？请阅读用户指南中的 [智能体如何记住你](../02-user-guide/07-memory/01-how-memory-works) +- 想了解普通遗忘机制？请阅读 [自动学习与遗忘](../02-user-guide/07-memory/04-auto-learning) - 想了解进化机制？请阅读 [自我进化](./self-evolution) From 648ceb5be42b253abe0fc5c7053762c9d0237496 Mon Sep 17 00:00:00 2001 From: yi-ge Date: Tue, 28 Jul 2026 11:06:54 +0800 Subject: [PATCH 2/5] =?UTF-8?q?docs(concepts):=20P0=20optimization=20?= =?UTF-8?q?=E2=80=94=20EN=20auto-dream,=20index=20sync,=20compute=20model,?= =?UTF-8?q?=20ASCII=E2=86=92SVG?= MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit - P0-2: Create EN 11-auto-dream.md (125 lines, full translation) - P0-3: Add Task Orchestration & Auto Dream rows to EN index.md - P0-4: Add 'Official Cloud Compute' and 'Generative vs. Chat Models' to EN 08-compute-model.md - P0-5: Replace ASCII art with SVG references in EN 07-three-layer-control.md --- .../04-concepts/07-three-layer-control.md | 25 +--- .../current/04-concepts/08-compute-model.md | 15 ++- .../current/04-concepts/11-auto-dream.md | 125 ++++++++++++++++++ .../current/04-concepts/index.md | 2 + 4 files changed, 142 insertions(+), 25 deletions(-) create mode 100644 i18n/en/docusaurus-plugin-content-docs/current/04-concepts/11-auto-dream.md diff --git a/i18n/en/docusaurus-plugin-content-docs/current/04-concepts/07-three-layer-control.md b/i18n/en/docusaurus-plugin-content-docs/current/04-concepts/07-three-layer-control.md index b206a48..e4868d4 100644 --- a/i18n/en/docusaurus-plugin-content-docs/current/04-concepts/07-three-layer-control.md +++ b/i18n/en/docusaurus-plugin-content-docs/current/04-concepts/07-three-layer-control.md @@ -18,20 +18,7 @@ DesireCore answers these questions through the **Three-Layer Controllability** m ## Three-Layer Controllability Model -``` - ┌───────────────┐ - │ L3: Reversible│ ← Can fix mistakes - │ Reversible │ - └───────┬───────┘ - ┌─────────┴─────────┐ - │ L2: Controllable│ ← Can decide what it does - │ Controllable │ - └─────────┬─────────┘ - ┌─────────────┴─────────────┐ - │ L1: Visible │ ← Can see what it's doing - │ Visible │ - └───────────────────────────┘ -``` +![Three-Layer Pyramid](/img/concepts/three-layer-pyramid.svg) ### L1: Visible — "I Can See What It's Doing" @@ -79,15 +66,7 @@ DesireCore answers these questions through the **Three-Layer Controllability** m These three layers form a **trust ladder**: -``` -Visible → Understand behavior → Establish preliminary trust - ↓ -Controllable → Set boundaries → Trust further enhanced - ↓ -Reversible → Eliminate worries → Complete trust - ↓ -Dare to delegate more tasks → Gain greater value -``` +![Trust Ladder](/img/concepts/trust-ladder.svg) All three layers are indispensable: diff --git a/i18n/en/docusaurus-plugin-content-docs/current/04-concepts/08-compute-model.md b/i18n/en/docusaurus-plugin-content-docs/current/04-concepts/08-compute-model.md index 094497d..082087d 100644 --- a/i18n/en/docusaurus-plugin-content-docs/current/04-concepts/08-compute-model.md +++ b/i18n/en/docusaurus-plugin-content-docs/current/04-concepts/08-compute-model.md @@ -103,11 +103,22 @@ Different tasks suit different models. DesireCore helps you automatically select | tts | Speech synthesis | TTS models | | asr | Speech recognition | ASR models | +## Official Cloud Compute + +If you don't want to register vendor accounts yourself, you can sign in to your DesireCore account to use official cloud compute. Cloud compute appears in the resource panel as the `desirecore-cloud` provider and syncs the models available to your current account. + +Cloud compute features: + +- Usage based on subscription or credits +- Model list synced via account API +- Model selector displays relative cost multiplier +- Can coexist with BYOK vendors + You can customize these mappings in settings — for example, point daily conversation to cheaper models, and important tasks to the strongest models. -## Pre-configured Compute Packages +## Generative vs. Chat Models -If you don't want to register vendor accounts yourself, DesireCore also plans to offer pre-configured compute packages — out-of-the-box, pay-as-you-go. But BYOK mode is always available; you won't be locked into any single vendor. +Image, video, music, TTS and other generative models do not appear as regular chat models. They are invoked by `GenerateImage`, `GenerateVideo`, TTS, or music generation skills. The chat model selector only shows models suitable for conversation, and filters out entries that have been explicitly verified as failed or deprecated. ## Next Steps diff --git a/i18n/en/docusaurus-plugin-content-docs/current/04-concepts/11-auto-dream.md b/i18n/en/docusaurus-plugin-content-docs/current/04-concepts/11-auto-dream.md new file mode 100644 index 0000000..dfe3efd --- /dev/null +++ b/i18n/en/docusaurus-plugin-content-docs/current/04-concepts/11-auto-dream.md @@ -0,0 +1,125 @@ +--- +title: Auto Dream — Lossless Forgetting +description: Understanding the Auto Dream mechanism — how agents "dream" to consolidate memories and preserve value through compression +keywords: [Auto Dream, Lossless Forgetting, memory consolidation, memory compression, knowledge distillation] +--- + +# Auto Dream — Lossless Forgetting + +## Why Do We Need "Dreaming" + +You may know that during deep sleep, the human brain isn't idle — it's organizing the day's experiences, weaving fragments into complete memories, discarding irrelevant details, and strengthening what truly matters. Neuroscientists call this "memory consolidation." + +DesireCore's agents face a similar challenge. As your interactions accumulate, the memory library continuously grows: + +- **Slower retrieval** — When memories grow from hundreds to tens of thousands, finding relevant memories means sifting through more "old notes" +- **Higher costs** — A larger memory library means more storage and compute overhead +- **Noise interference** — A flood of trivial, scattered memories can drown out truly important information + +But brutally deleting old memories would discard valuable experience and patterns — the things you spent months "teaching" the agent shouldn't just vanish. + +**Auto Dream is the best balance between the two**: instead of deleting, it consolidates scattered memories into more refined knowledge, just like dreaming. + +## What is Auto Dream + +Auto Dream is a **memory consolidation mechanism** that runs periodically in the background. Its approach isn't simple compression or deletion — instead, it: + +1. Identifies thematically related scattered memories +2. Extracts common patterns and core information +3. Restructures multiple fragmented memories into organized knowledge + +For example: you have a stack of sticky notes that read "user likes short sentences," "user dislikes exclamation marks," "user prefers active voice," "user requires paragraphs under three lines." Auto Dream takes these four notes and merges them into a clear "Writing Style Guide" — no information is lost, but the form is more refined and useful. + +### How It Differs from Ordinary Forgetting + +| | Ordinary Forgetting (Compression/Cleanup) | Auto Dream (Lossless Forgetting) | +|---|---|---| +| **How it works** | Reduces precision or deletes directly | Extracts essence, then reorganizes | +| **Information loss** | Lossy — details are unrecoverable | Lossless — semantics fully preserved | +| **Trigger condition** | Storage limit / memory age | Periodic automatic trigger | +| **Output** | Compressed summary or empty | Structured knowledge entry | +| **Traceability** | Original text unrecoverable after compression | Original memories retained in version history | + +In short, ordinary forgetting is "losing things," while Auto Dream is "organizing the bookshelf" — the books are still there, just arranged more neatly. + +## How Auto Dream Works + +### The Consolidation Process + +Auto Dream's work proceeds in four phases: + +**Scan** — The system periodically scans memories in "archived" status, identifying groups of thematically similar memories that can be merged. For example, 20 scattered memories about your writing habits would be grouped together. + +**Associate** — Within each group, the system analyzes semantic relationships between memories: which are complementary? Which are duplicates? Which contain contradictions where the newer one should take precedence? This step is similar to how the brain builds associations between memories during dreaming. + +**Distill** — The system extracts core patterns and key information from the associated memory group, removing redundancy and noise. It doesn't delete details — it concentrates information scattered across multiple memories into a more refined expression. + +**Reorganize** — The distilled information is reorganized into a structured knowledge entry and written to the memory library. The original scattered memories are not deleted — they're marked as "consolidated" and retained in version history for future reference. + +### Memory Lifecycle (Including the Dream Phase) + +In [Auto-Learning and Forgetting](/docs/02-user-guide/07-memory/04-auto-learning), we introduced the five-stage memory lifecycle. With Auto Dream, the complete lifecycle becomes: + +``` +Active → Recent → Archived → 💤 Dream → Compressed → Cleaned +``` + +The Dream phase is inserted **between Archived and Compressed**. When a group of archived memories meets the Dream criteria, the system attempts consolidation first, rather than compressing directly. After successful consolidation, the resulting structured knowledge re-enters the lifecycle as "Recent" status — because it's newly generated, valuable knowledge that deserves higher priority. + +## Dream Output Example + +Suppose your writing advisor agent has accumulated these scattered memories from conversations over the past month: + +> - "User wants articles to get straight to the point, no preamble" +> - "User prefers short paragraphs, no more than 3 sentences each" +> - "User doesn't like passive voice" +> - "User requires headings to start with verbs" +> - "User mentioned the target audience is technical managers" +> - "User wants a call-to-action at the end" +> - ... along with a dozen other similar fragments + +After Auto Dream consolidation, these memories are distilled into a structured **Writing Style Profile**: + +> **User Writing Style Profile** +> +> - **Target audience**: Technical managers +> - **Overall style**: Concise and direct, conclusions over preambles +> - **Paragraph structure**: Short paragraphs (≤3 sentences), headings use verbs +> - **Voice preference**: Active voice preferred, avoid passive constructions +> - **Article structure**: Straight to the point at the start, call-to-action at the end + +20 scattered memories become 1 clear profile. The information volume hasn't decreased, but retrieval efficiency and usability are dramatically improved — the agent only needs to reference this one profile next time it helps you write, instead of sifting through 20 sticky notes. + +## Relationship with Self-Evolution + +Auto Dream isn't just a memory management tool — it's also essential nourishment for the agent's [Self-Evolution](/docs/concepts/self-evolution). + +When Dream consolidates scattered experiences into structured knowledge, the evolution mechanism can further leverage these results: + +- **Rule extraction** — If a pattern in the profile is stable enough (e.g., "user consistently requires short paragraphs"), the evolution mechanism may suggest elevating it to a formal behavioral rule +- **Skill optimization** — Consolidated knowledge helps the agent adjust existing skill execution strategies +- **Contradiction discovery** — If Dream finds contradictions between memories (e.g., the user early on said "be detailed" but later said "be concise"), it flags them for the evolution mechanism to handle + +You could say Auto Dream is the "preprocessor" for evolution — it processes raw materials into semi-finished products, allowing the evolution mechanism to more efficiently transform them into genuine capability improvements. + +## User Control + +Auto Dream is completely under your control. You can: + +| What You Want to Do | How to Do It | +|---------------------|-------------| +| View Dream logs | Switch to the "Dream Records" tab in memory management | +| Manually trigger Dream | Click "Organize Memories" and select "Deep Consolidation" mode | +| Protect memories from consolidation | Mark memories as "Pinned" — pinned memories don't participate in Dream | +| Adjust Dream frequency | Modify the automatic consolidation trigger interval in settings | +| Trace back to original memories | Click "View Sources" in Dream-produced knowledge entries to expand all original memories | + +:::tip Let It Dream Freely +Auto Dream never deletes any original memory. Even if you're not satisfied with the consolidation result, you can always view and restore the original scattered memories. +::: + +## Next Steps + +- Want to understand how memory works end-to-end? Read [How the Agent Remembers You](/docs/02-user-guide/07-memory/01-how-memory-works) in the User Guide +- Want to understand the ordinary forgetting mechanism? Read [Auto-Learning and Forgetting](/docs/02-user-guide/07-memory/04-auto-learning) +- Want to understand the evolution mechanism? Read [Self-Evolution](./self-evolution) diff --git a/i18n/en/docusaurus-plugin-content-docs/current/04-concepts/index.md b/i18n/en/docusaurus-plugin-content-docs/current/04-concepts/index.md index fdffa19..98baee3 100644 --- a/i18n/en/docusaurus-plugin-content-docs/current/04-concepts/index.md +++ b/i18n/en/docusaurus-plugin-content-docs/current/04-concepts/index.md @@ -56,6 +56,8 @@ DesireCore is built around the core philosophy of "cultivating and hosting digit | [Self-Evolution](./self-evolution) | The secret to understanding you better over time | Want to understand agent growth | | [Three-Layer Controllability](./three-layer-control) | Visible, Controllable, Reversible — you always hold the initiative | Want to understand safety guarantees | | [Compute Model](./compute-model) | Bring your own keys, choose your own engines | Want to understand AI model configuration | +| [Task Orchestration](./task-orchestration) | Automatic task decomposition and best-agent matching — your "project manager" | Want to understand task distribution | +| [Auto Dream](./auto-dream) | Lossless forgetting — preserving value through compression | Want to understand memory consolidation | ## Reading Recommendations From ce0cce82c21e29b2e3fe0379d1ca89bf5f4d3f65 Mon Sep 17 00:00:00 2001 From: yi-ge Date: Tue, 28 Jul 2026 11:15:43 +0800 Subject: [PATCH 3/5] fix: correct broken links to use URL paths without numeric prefixes --- docs/04-concepts/10-task-orchestration.md | 2 +- docs/04-concepts/11-auto-dream.md | 6 +++--- 2 files changed, 4 insertions(+), 4 deletions(-) diff --git a/docs/04-concepts/10-task-orchestration.md b/docs/04-concepts/10-task-orchestration.md index fb61dc7..a5f8977 100644 --- a/docs/04-concepts/10-task-orchestration.md +++ b/docs/04-concepts/10-task-orchestration.md @@ -120,5 +120,5 @@ DesireCore 的**智能任务编排引擎**就是你的"项目经理"。当你提 ## 下一步 - 想了解步骤的可控性？请阅读 [固化/灵活/人闸门](./step-types) -- 想了解多智能体如何协作？请阅读用户指南中的 [跨智能体协作](../02-user-guide/04-delegation/06-cross-agent) +- 想了解多智能体如何协作？请阅读用户指南中的 [跨智能体协作](../user-guide/delegation/cross-agent) - 想了解任务的回执系统？请阅读 [回执系统](./receipt-system) diff --git a/docs/04-concepts/11-auto-dream.md b/docs/04-concepts/11-auto-dream.md index 6c8c1fe..db4c99c 100644 --- a/docs/04-concepts/11-auto-dream.md +++ b/docs/04-concepts/11-auto-dream.md @@ -58,7 +58,7 @@ Auto Dream 的工作分为四个阶段： ### 记忆生命周期（含 Dream 阶段） -在 [自动学习与遗忘](../02-user-guide/07-memory/04-auto-learning) 中，我们介绍了记忆的五阶段生命周期。加入 Auto Dream 后，完整的生命周期变为： +在 [自动学习与遗忘](../user-guide/memory/auto-learning) 中，我们介绍了记忆的五阶段生命周期。加入 Auto Dream 后，完整的生命周期变为： ``` 活跃 → 近期 → 归档 → 💤 Dream → 压缩 → 清理 @@ -120,6 +120,6 @@ Auto Dream 不会删除任何原始记忆。即使你对整合结果不满意， ## 下一步 -- 想了解记忆的完整工作原理？请阅读用户指南中的 [智能体如何记住你](../02-user-guide/07-memory/01-how-memory-works) -- 想了解普通遗忘机制？请阅读 [自动学习与遗忘](../02-user-guide/07-memory/04-auto-learning) +- 想了解记忆的完整工作原理？请阅读用户指南中的 [智能体如何记住你](../user-guide/memory/how-memory-works) +- 想了解普通遗忘机制？请阅读 [自动学习与遗忘](../user-guide/memory/auto-learning) - 想了解进化机制？请阅读 [自我进化](./self-evolution) From 17244826e3cb43b1acb84c5d5e50b5dd0fd2778f Mon Sep 17 00:00:00 2001 From: yi-ge Date: Tue, 28 Jul 2026 11:35:51 +0800 Subject: [PATCH 4/5] fix: correct broken links in EN concepts pages --- .../current/04-concepts/10-task-orchestration.md | 2 +- .../current/04-concepts/11-auto-dream.md | 6 +++--- 2 files changed, 4 insertions(+), 4 deletions(-) diff --git a/i18n/en/docusaurus-plugin-content-docs/current/04-concepts/10-task-orchestration.md b/i18n/en/docusaurus-plugin-content-docs/current/04-concepts/10-task-orchestration.md index 48c403d..95a644f 100644 --- a/i18n/en/docusaurus-plugin-content-docs/current/04-concepts/10-task-orchestration.md +++ b/i18n/en/docusaurus-plugin-content-docs/current/04-concepts/10-task-orchestration.md @@ -120,5 +120,5 @@ After the whole process finishes, the consolidated receipt the engine produces i ## Next Steps - Want to understand step controllability? Read [Hardened/Flexible/Human Gate](./step-types) -- Want to understand how multiple agents collaborate? Read [Cross-Agent Collaboration](../02-user-guide/04-delegation/06-cross-agent.md) in the user guide +- Want to understand how multiple agents collaborate? Read [Cross-Agent Collaboration](../user-guide/delegation/cross-agent) in the user guide - Want to understand the task receipt system? Read [Receipt System](./receipt-system) diff --git a/i18n/en/docusaurus-plugin-content-docs/current/04-concepts/11-auto-dream.md b/i18n/en/docusaurus-plugin-content-docs/current/04-concepts/11-auto-dream.md index dfe3efd..6ff86b1 100644 --- a/i18n/en/docusaurus-plugin-content-docs/current/04-concepts/11-auto-dream.md +++ b/i18n/en/docusaurus-plugin-content-docs/current/04-concepts/11-auto-dream.md @@ -58,7 +58,7 @@ Auto Dream's work proceeds in four phases: ### Memory Lifecycle (Including the Dream Phase) -In [Auto-Learning and Forgetting](/docs/02-user-guide/07-memory/04-auto-learning), we introduced the five-stage memory lifecycle. With Auto Dream, the complete lifecycle becomes: +In [Auto-Learning and Forgetting](/docs/user-guide/memory/auto-learning), we introduced the five-stage memory lifecycle. With Auto Dream, the complete lifecycle becomes: ``` Active → Recent → Archived → 💤 Dream → Compressed → Cleaned @@ -120,6 +120,6 @@ Auto Dream never deletes any original memory. Even if you're not satisfied with ## Next Steps -- Want to understand how memory works end-to-end? Read [How the Agent Remembers You](/docs/02-user-guide/07-memory/01-how-memory-works) in the User Guide -- Want to understand the ordinary forgetting mechanism? Read [Auto-Learning and Forgetting](/docs/02-user-guide/07-memory/04-auto-learning) +- Want to understand how memory works end-to-end? Read [How the Agent Remembers You](/docs/user-guide/memory/how-memory-works) in the User Guide +- Want to understand the ordinary forgetting mechanism? Read [Auto-Learning and Forgetting](/docs/user-guide/memory/auto-learning) - Want to understand the evolution mechanism? Read [Self-Evolution](./self-evolution) From 41387569cfd7a24b91f048fa442622845eb5166a Mon Sep 17 00:00:00 2001 From: yi-ge Date: Tue, 28 Jul 2026 11:43:21 +0800 Subject: [PATCH 5/5] fix: use relative paths in EN auto-dream links --- .../current/04-concepts/11-auto-dream.md | 8 ++++---- 1 file changed, 4 insertions(+), 4 deletions(-) diff --git a/i18n/en/docusaurus-plugin-content-docs/current/04-concepts/11-auto-dream.md b/i18n/en/docusaurus-plugin-content-docs/current/04-concepts/11-auto-dream.md index 6ff86b1..6d83a6a 100644 --- a/i18n/en/docusaurus-plugin-content-docs/current/04-concepts/11-auto-dream.md +++ b/i18n/en/docusaurus-plugin-content-docs/current/04-concepts/11-auto-dream.md @@ -58,7 +58,7 @@ Auto Dream's work proceeds in four phases: ### Memory Lifecycle (Including the Dream Phase) -In [Auto-Learning and Forgetting](/docs/user-guide/memory/auto-learning), we introduced the five-stage memory lifecycle. With Auto Dream, the complete lifecycle becomes: +In [Auto-Learning and Forgetting](../user-guide/memory/auto-learning), we introduced the five-stage memory lifecycle. With Auto Dream, the complete lifecycle becomes: ``` Active → Recent → Archived → 💤 Dream → Compressed → Cleaned @@ -92,7 +92,7 @@ After Auto Dream consolidation, these memories are distilled into a structured * ## Relationship with Self-Evolution -Auto Dream isn't just a memory management tool — it's also essential nourishment for the agent's [Self-Evolution](/docs/concepts/self-evolution). +Auto Dream isn't just a memory management tool — it's also essential nourishment for the agent's [Self-Evolution](./self-evolution). When Dream consolidates scattered experiences into structured knowledge, the evolution mechanism can further leverage these results: @@ -120,6 +120,6 @@ Auto Dream never deletes any original memory. Even if you're not satisfied with ## Next Steps -- Want to understand how memory works end-to-end? Read [How the Agent Remembers You](/docs/user-guide/memory/how-memory-works) in the User Guide -- Want to understand the ordinary forgetting mechanism? Read [Auto-Learning and Forgetting](/docs/user-guide/memory/auto-learning) +- Want to understand how memory works end-to-end? Read [How the Agent Remembers You](../user-guide/memory/how-memory-works) in the User Guide +- Want to understand the ordinary forgetting mechanism? Read [Auto-Learning and Forgetting](../user-guide/memory/auto-learning) - Want to understand the evolution mechanism? Read [Self-Evolution](./self-evolution)