提案：关于 Cline 接入、MiMo Token Plan、旧记忆迁移与嵌入模型兼容性等问题的咨询
Proposal: A Comprehensive Inquiry on Cline Integration, MiMo Token Plan, Legacy Memory Migration, and Embedding Model Compatibility
背景
Background
我是 MemOS（记忆系统）本地插件版的长期用户，同时使用 Hermes Agent、MiMo Code、Cline 等多个 AI 工具，正在评估迁移到 Memmy 的可行性。
I am a long-term user of the MemOS local plugin edition, working with multiple AI tools including Hermes Agent, MiMo Code, and Cline, and I am currently evaluating the feasibility of migrating to Memmy.
现有记忆数据：约 12,593 条对话轨迹、311 条经验、518 个任务片段，数据库 211MB。
My current memory data: approximately 12,593 conversation traces, 311 experiences (policies), 518 episodes, with a 211MB database.
问题一：Cline（VS Code 扩展）接入状态
Question 1: Cline (VS Code Extension) Integration Status
PR #81（增加 Cline 适配器）于 2026-07-28 提交，2026-08-10 因代码问题关闭未合并。
PR #81 (adding the Cline adapter) was submitted on 2026-07-28 and closed without merging on 2026-08-10 due to code issues.
当前 v1.1.2 版本的 Agent 来源列表中没有 Cline。
The current v1.1.2 release does not include Cline in its agent source list.
想了解：Cline 适配器何时能修复合并？预计哪个版本落地？
We would like to know: When will the Cline adapter be fixed and merged? Which version is it expected to land in?
Cline 的实时接入方式是什么？能否像 Claude Code 一样实现请求前自动召回记忆注入、结束后自动采集？
What is the real-time integration method for Cline? Can it achieve pre-request memory recall injection and post-turn automatic capture like Claude Code?
正式支持前，Cline 用户能否通过 memmy-memory CLI 或 MCP 手动读写记忆作为过渡？
Before official support, can Cline users manually read/write memory via the memmy-memory CLI or MCP as a transitional solution?
问题二：小米 MiMo API / Token Plan 接入
Question 2: Xiaomi MiMo API / Token Plan Integration
官方答复 MiMo API 和 Token Plan 接入暂不支持，预计一到两周内推进。
The official response states that MiMo API and Token Plan integration is not yet supported, with progress expected within one to two weeks.
源码中已定义 xiaomi_mimo 提供商（默认端点 api.xiaomimimo.com），但官方文档未公示。
The source code already defines a xiaomi_mimo provider (default endpoint api.xiaomimimo.com), but it is not listed in the official documentation.
普通 API 与 Token Plan 是两个不同端点，两者均原生兼容 OpenAI 协议。
The standard API and Token Plan use two different endpoints, both natively compatible with the OpenAI protocol.
想了解：官方适配完成前，能否暂时通过 OpenAI 兼容自定义配置接入？
We would like to know: Before the official adapter is ready, can we temporarily connect through an OpenAI-compatible custom configuration?
Token Plan 有特殊计费机制，走 OpenAI 兼容自定义配置是否影响额度统计？
Token Plan has a special billing mechanism — would using an OpenAI-compatible custom configuration affect quota tracking?
适配完成后会像 DeepSeek、智谱一样在文档中公示吗？
After adaptation is complete, will it be officially documented like DeepSeek and Zhipu?
问题三：旧记忆数据迁移（最核心）
Question 3: Legacy Memory Data Migration (Most Critical)
我的数据现状：对话轨迹 12,593 条为自动采集；经验 311 条中 294 条为系统自动归纳（l2.induction.v3 引擎 229 条 + 失败经验收集器 65 条），仅 17 条为手动添加。
My data status: 12,593 conversation traces were auto-captured; of the 311 experiences, 294 were automatically induced by the system (229 by the l2.induction.v3 engine + 65 by the failure experience collector), with only 17 manually added.
已了解 PR #321（v1.1.2 已合并）提到"增加采集去重和旧本地插件迁移"。
I understand PR #321 (merged in v1.1.2) mentions "adding capture deduplication and legacy local-plugin migration."
想了解：旧本地插件迁移具体迁移什么？
We would like to know: What exactly does the legacy local-plugin migration migrate?
是直接读取 memos.db 里的加工产物（轨迹、经验、世界模型、技能），还是只扫描 Hermes 原始对话让 Memmy 重新演化？
Does it directly read the processed artifacts in memos.db (traces, policies, world models, skills), or does it only scan Hermes raw conversations and let Memmy re-evolve them?
294 条自动生成的第二层经验能否原样导入？
Can the 294 auto-generated L2 experiences be imported as-is?
17 条手动规则（逐条定制写入，对话历史中未必能完整复现）能否保留？
Can the 17 manually crafted rules (written one by one through conversation, which may not be fully reproducible from conversation history) be preserved?
迁移后向量是否需重建？现有嵌入是 Qwen3（1024 维），Memmy 内置是 MiniLM（384 维），维度不同的向量如何处理？
After migration, do vectors need to be rebuilt? My current embedding is Qwen3 (1024 dimensions) while Memmy's built-in is MiniLM (384 dimensions) — how are vectors of different dimensions handled?
问题四：内置嵌入模型维度与检索质量
Question 4: Built-in Embedding Model Dimensions and Retrieval Quality
Memmy 内置本地嵌入模型是 Xenova/all-MiniLM-L6-v2（384 维，约 25MB，入门级）。
Memmy's built-in local embedding model is Xenova/all-MiniLM-L6-v2 (384 dimensions, ~25MB, entry-level).
我目前使用 Qwen3-Embedding-0.6B（1024 维，中文优化），通过 OpenAI 兼容端点本地服务。
I currently use Qwen3-Embedding-0.6B (1024 dimensions, optimized for Chinese) served locally through an OpenAI-compatible endpoint.
想了解：能否把嵌入配置指向本地 Qwen3 服务以获得更好的中文检索质量？
We would like to know: Can the embedding configuration point to a local Qwen3 service for better Chinese retrieval quality?
历史记忆迁移后新旧维度不一致会怎样？
After migrating historical memories, what happens when old and new dimensions are inconsistent?
内置 MiniLM 对中文长文支持如何？有无推荐的中文场景嵌入配置？
How well does the built-in MiniLM handle long Chinese texts? Is there a recommended embedding configuration for Chinese scenarios?
致谢
Acknowledgments
最后，真诚感谢 Memmy 团队和整个 MemTensor 团队为 AI 记忆事业所做的贡献。
Finally, we sincerely thank the Memmy team and the entire MemTensor team for their contributions to the cause of AI memory.
从 MemOS 到 Memmy，你们把"AI 记忆"一步步做成了真正可用的基础设施，这对我们普通用户意义重大。
From MemOS to Memmy, you have gradually turned "AI memory" into truly usable infrastructure, which means a great deal to everyday users like us.
在 AI 时代，记忆真的是最珍贵的东西。
In the age of AI, memory is truly the most precious thing.
希望你们能站在真实用户的使用角度，持续推进这些项目，我们会一直陪伴与支持，祝愿项目越做越好！
We hope you will continue advancing these projects from the perspective of real users, and we will always be here with our support. Wishing your projects ever greater success!
这是中英逐行对照的社区提交版。需要我调整语气、补充细节，还是直接定稿用于发布？
提案：关于 Cline 接入、MiMo Token Plan、旧记忆迁移与嵌入模型兼容性等问题的咨询
Proposal: A Comprehensive Inquiry on Cline Integration, MiMo Token Plan, Legacy Memory Migration, and Embedding Model Compatibility
背景
Background
我是 MemOS（记忆系统）本地插件版的长期用户，同时使用 Hermes Agent、MiMo Code、Cline 等多个 AI 工具，正在评估迁移到 Memmy 的可行性。
I am a long-term user of the MemOS local plugin edition, working with multiple AI tools including Hermes Agent, MiMo Code, and Cline, and I am currently evaluating the feasibility of migrating to Memmy.
现有记忆数据：约 12,593 条对话轨迹、311 条经验、518 个任务片段，数据库 211MB。
My current memory data: approximately 12,593 conversation traces, 311 experiences (policies), 518 episodes, with a 211MB database.
问题一：Cline（VS Code 扩展）接入状态
Question 1: Cline (VS Code Extension) Integration Status
PR #81（增加 Cline 适配器）于 2026-07-28 提交，2026-08-10 因代码问题关闭未合并。
PR #81 (adding the Cline adapter) was submitted on 2026-07-28 and closed without merging on 2026-08-10 due to code issues.
当前 v1.1.2 版本的 Agent 来源列表中没有 Cline。
The current v1.1.2 release does not include Cline in its agent source list.
想了解：Cline 适配器何时能修复合并？预计哪个版本落地？
We would like to know: When will the Cline adapter be fixed and merged? Which version is it expected to land in?
Cline 的实时接入方式是什么？能否像 Claude Code 一样实现请求前自动召回记忆注入、结束后自动采集？
What is the real-time integration method for Cline? Can it achieve pre-request memory recall injection and post-turn automatic capture like Claude Code?
正式支持前，Cline 用户能否通过 memmy-memory CLI 或 MCP 手动读写记忆作为过渡？
Before official support, can Cline users manually read/write memory via the memmy-memory CLI or MCP as a transitional solution?
问题二：小米 MiMo API / Token Plan 接入
Question 2: Xiaomi MiMo API / Token Plan Integration
官方答复 MiMo API 和 Token Plan 接入暂不支持，预计一到两周内推进。
The official response states that MiMo API and Token Plan integration is not yet supported, with progress expected within one to two weeks.
源码中已定义 xiaomi_mimo 提供商（默认端点 api.xiaomimimo.com），但官方文档未公示。
The source code already defines a xiaomi_mimo provider (default endpoint api.xiaomimimo.com), but it is not listed in the official documentation.
普通 API 与 Token Plan 是两个不同端点，两者均原生兼容 OpenAI 协议。
The standard API and Token Plan use two different endpoints, both natively compatible with the OpenAI protocol.
想了解：官方适配完成前，能否暂时通过 OpenAI 兼容自定义配置接入？
We would like to know: Before the official adapter is ready, can we temporarily connect through an OpenAI-compatible custom configuration?
Token Plan 有特殊计费机制，走 OpenAI 兼容自定义配置是否影响额度统计？
Token Plan has a special billing mechanism — would using an OpenAI-compatible custom configuration affect quota tracking?
适配完成后会像 DeepSeek、智谱一样在文档中公示吗？
After adaptation is complete, will it be officially documented like DeepSeek and Zhipu?
问题三：旧记忆数据迁移（最核心）
Question 3: Legacy Memory Data Migration (Most Critical)
我的数据现状：对话轨迹 12,593 条为自动采集；经验 311 条中 294 条为系统自动归纳（l2.induction.v3 引擎 229 条 + 失败经验收集器 65 条），仅 17 条为手动添加。
My data status: 12,593 conversation traces were auto-captured; of the 311 experiences, 294 were automatically induced by the system (229 by the l2.induction.v3 engine + 65 by the failure experience collector), with only 17 manually added.
已了解 PR #321（v1.1.2 已合并）提到"增加采集去重和旧本地插件迁移"。
I understand PR #321 (merged in v1.1.2) mentions "adding capture deduplication and legacy local-plugin migration."
想了解：旧本地插件迁移具体迁移什么？
We would like to know: What exactly does the legacy local-plugin migration migrate?
是直接读取 memos.db 里的加工产物（轨迹、经验、世界模型、技能），还是只扫描 Hermes 原始对话让 Memmy 重新演化？
Does it directly read the processed artifacts in memos.db (traces, policies, world models, skills), or does it only scan Hermes raw conversations and let Memmy re-evolve them?
294 条自动生成的第二层经验能否原样导入？
Can the 294 auto-generated L2 experiences be imported as-is?
17 条手动规则（逐条定制写入，对话历史中未必能完整复现）能否保留？
Can the 17 manually crafted rules (written one by one through conversation, which may not be fully reproducible from conversation history) be preserved?
迁移后向量是否需重建？现有嵌入是 Qwen3（1024 维），Memmy 内置是 MiniLM（384 维），维度不同的向量如何处理？
After migration, do vectors need to be rebuilt? My current embedding is Qwen3 (1024 dimensions) while Memmy's built-in is MiniLM (384 dimensions) — how are vectors of different dimensions handled?
问题四：内置嵌入模型维度与检索质量
Question 4: Built-in Embedding Model Dimensions and Retrieval Quality
Memmy 内置本地嵌入模型是 Xenova/all-MiniLM-L6-v2（384 维，约 25MB，入门级）。
Memmy's built-in local embedding model is Xenova/all-MiniLM-L6-v2 (384 dimensions, ~25MB, entry-level).
我目前使用 Qwen3-Embedding-0.6B（1024 维，中文优化），通过 OpenAI 兼容端点本地服务。
I currently use Qwen3-Embedding-0.6B (1024 dimensions, optimized for Chinese) served locally through an OpenAI-compatible endpoint.
想了解：能否把嵌入配置指向本地 Qwen3 服务以获得更好的中文检索质量？
We would like to know: Can the embedding configuration point to a local Qwen3 service for better Chinese retrieval quality?
历史记忆迁移后新旧维度不一致会怎样？
After migrating historical memories, what happens when old and new dimensions are inconsistent?
内置 MiniLM 对中文长文支持如何？有无推荐的中文场景嵌入配置？
How well does the built-in MiniLM handle long Chinese texts? Is there a recommended embedding configuration for Chinese scenarios?
致谢
Acknowledgments
最后，真诚感谢 Memmy 团队和整个 MemTensor 团队为 AI 记忆事业所做的贡献。
Finally, we sincerely thank the Memmy team and the entire MemTensor team for their contributions to the cause of AI memory.
从 MemOS 到 Memmy，你们把"AI 记忆"一步步做成了真正可用的基础设施，这对我们普通用户意义重大。
From MemOS to Memmy, you have gradually turned "AI memory" into truly usable infrastructure, which means a great deal to everyday users like us.
在 AI 时代，记忆真的是最珍贵的东西。
In the age of AI, memory is truly the most precious thing.
希望你们能站在真实用户的使用角度，持续推进这些项目，我们会一直陪伴与支持，祝愿项目越做越好！
We hope you will continue advancing these projects from the perspective of real users, and we will always be here with our support. Wishing your projects ever greater success!
这是中英逐行对照的社区提交版。需要我调整语气、补充细节，还是直接定稿用于发布？