优化记忆提取

Author: linyuan.yang

packages/admin/src/i18n/en.ts

@@ -533,6 +533,8 @@ export default {
     name_placeholder: 'e.g. Default Memory',
     writer_model: 'Writer Model',
     writer_model_placeholder: 'MemoryLLM (reasoning-capable recommended)',
+    selector_model: 'Selector Model',
+    use_writer_model: 'Use Writer Model',
     writer_prompt: 'Writer System Prompt',
     read_prompt: 'Read Path Template',
     run_consolidate: 'Clean Up Memories',

packages/admin/src/i18n/zh.ts

@@ -533,6 +533,8 @@ export default {
     name_placeholder: '如：默认长期记忆',
     writer_model: 'Writer 模型',
     writer_model_placeholder: '后台 MemoryLLM（建议 reasoning 强的模型）',
+    selector_model: 'Selector 模型',
+    use_writer_model: '使用 Writer 模型',
     writer_prompt: 'Writer 系统 Prompt',
     read_prompt: 'Read 路径模板',
     run_consolidate: '整理记忆',

packages/admin/src/views/automation/MemoryProfilesView.vue

@@ -12,6 +12,7 @@ interface MemoryProfileForm {
   name: string
   enabled: boolean
   writerModel: string
+  selectorModel: string
   writerPromptFile: string
   readPromptFile: string
 }
@@ -42,6 +43,7 @@ const columns = computed<STableColumn[]>(() => [
   { key: 'name',         label: t('common.name'),                primary: true },
   { key: 'enabled',      label: t('common.enabled'),             width: '100px', align: 'center' },
   { key: 'writerModel',  label: t('memory_profiles.writer_model'), width: '200px' },
+  { key: 'selectorModel', label: t('memory_profiles.selector_model'), width: '200px' },
   { key: 'ops',          label: t('common.ops'),                 ops: true, width: '320px', align: 'center' },
 ])
 
@@ -55,6 +57,7 @@ function emptyForm(): MemoryProfileForm {
     name: '',
     enabled: true,
     writerModel: '',
+    selectorModel: '',
     writerPromptFile: '',
     readPromptFile: '',
   }
@@ -94,6 +97,7 @@ function openEdit(id: string) {
     name: p.name || '',
     enabled: !!p.enabled,
     writerModel: p.writerModel || '',
+    selectorModel: p.selectorModel || '',
     writerPromptFile: p.writerPromptFile || '',
     readPromptFile: p.readPromptFile || '',
   }
@@ -112,6 +116,7 @@ async function save() {
       enabled: form.value.enabled,
       writerModel: form.value.writerModel,
     }
+    if (form.value.selectorModel) body.selectorModel = form.value.selectorModel
     if (form.value.writerPromptFile) body.writerPromptFile = form.value.writerPromptFile
     if (form.value.readPromptFile)   body.readPromptFile   = form.value.readPromptFile
     const id = editingId.value
@@ -198,6 +203,7 @@ function modelLabel(id: string | undefined | null): string {
           </SBadge>
         </template>
         <template #writerModel="{ row }">{{ modelLabel(row.writerModel) }}</template>
+        <template #selectorModel="{ row }">{{ row.selectorModel ? modelLabel(row.selectorModel) : t('memory_profiles.use_writer_model') }}</template>
         <template #ops="{ row }">
           <div class="ops-row">
             <SButton type="outline" size="sm" @click="openMemoryViewer(row.id)">{{ t('memory_profiles.view_memories') }}</SButton>
@@ -230,6 +236,12 @@ function modelLabel(id: string | undefined | null): string {
           <option v-for="m in modelOptions" :key="m.id" :value="m.id">{{ m.label }}</option>
         </SSelect>
       </SFormItem>
+      <SFormItem :label="t('memory_profiles.selector_model')">
+        <SSelect v-model="form.selectorModel">
+          <option value="">{{ t('memory_profiles.use_writer_model') }}</option>
+          <option v-for="m in modelOptions" :key="m.id" :value="m.id">{{ m.label }}</option>
+        </SSelect>
+      </SFormItem>
       <SFormItem :label="t('memory_profiles.writer_prompt')">
         <SSelect v-model="form.writerPromptFile">
           <option value="">{{ t('common.default') }}</option>

packages/admin/src/views/models/ModelsView.vue

@@ -32,7 +32,7 @@ const refs = makeResourceRefs({
   profile: (p, id) => p.intentModel === id,
   session: (s, id) => s.intentModel === id,
   agent: (a, id) => a.model === id || a.compactModel === id,
-  memoryProfile: (p, id) => p.writerModel === id,
+  memoryProfile: (p, id) => p.writerModel === id || p.selectorModel === id,
   agendaProfile: (p, id) => p.syncModel === id,
 })
 const expandedIds = ref<string[]>([])

packages/docs-site/guide/channels.md

@@ -128,10 +128,16 @@ Talk to your XiaoAi smart speaker as a sbot channel.
 | 登录方式 | `passToken` (recommended, obtained via **sbox**), or Mi account password |
 | 密码 / passToken | The credential matching the selected login mode |
 | 登录设备 ID | Optional PassportSDK `deviceId`, obtained via **sbox** |
-| 音箱名称 | Target speaker name (matches a device in your Mi account) |
+| 音箱名称 | Target speaker name (matches a device in your Mi account). One per line to bind several speakers to this channel |
 
 2. The bot logs in to Mi Cloud, listens for voice messages on the device, and replies via TTS through the speaker
 
+A single channel can drive multiple speakers — list them all in 音箱名称 rather than creating one channel per speaker. The speakers share one Mi Cloud login, and each gets its own session, so a reply always goes back to the speaker that was spoken to. Creating several channels for the *same* Mi account does not work: each login invalidates the previous `serviceToken`, so all but the last channel fail with HTTP 401.
+
+::: warning
+Every speaker is polled independently, so N speakers means N × (1000 / 轮询间隔) requests per second against Mi Cloud. Raise 轮询间隔 when binding more than a few.
+:::
+
 ::: tip Getting `passToken` and `deviceId` with sbox
 Mi accounts often reject server-side password login (risk control / 2FA), so the recommended path is to log in once with **sbox** — a desktop helper toolbox — and copy the credentials it extracts:
 

packages/docs-site/guide/memory.md

@@ -2,20 +2,21 @@
 
 Sidebar → **Memory Profiles** (under **Tasks**), then enable it per-agent in the agent edit page → **Memory**.
 
-Memory is the agent's automatic long-term memory. A background **MemoryLLM** reviews each conversation after it goes idle and distills durable knowledge — user preferences, project facts, decisions, lessons learned — into memory entries. On later turns the agent reads them back via the `search_memory` and `read_memory` tools.
+Memory is the agent's automatic long-term memory. After each conversation, a background Selector reviews the full memory catalog for relevant entries, then the **MemoryLLM** reads those entries' complete bodies and applies final changes. On later turns the agent reads memories back via the `search_memory` and `read_memory` tools.
 
 Think of it as the agent learning from every conversation without you having to teach it explicitly — and without bloating the system prompt.
 
 ## How It Works
 
-1. **Extract** — after a conversation idles, the **writer model** silently reviews the exchange and writes new memories (or updates/removes existing ones).
-2. **Read** — on subsequent turns the agent calls:
-   - `search_memory` — fuzzy/keyword/semantic lookup across stored memories
+1. **Select** — after a conversation idles, the **selector model** matches the transcript against self-contained titles from every global + current-workspace entry. It uses one request while the title catalog fits its budget; for oversized catalogs it compresses the transcript once and matches titles in batches. The Selector never reads bodies. If no selector is configured, the writer model performs this step.
+2. **Extract** — the **writer model** reads the selected entries' complete bodies together with the conversation, then creates, updates, removes, or leaves memories unchanged.
+3. **Read** — on subsequent turns the agent calls:
+   - `search_memory` — BM25 keyword lookup across stored memories
    - `read_memory` — fetch the full body of a memory by its slug
-3. **Maintain** — background jobs keep the store healthy:
+4. **Maintain** — background jobs keep the store healthy:
    - **Consolidate** — merges and de-duplicates related memories
    - **Reconcile** — re-indexes and prunes stale entries
-4. **Delete** — removed memories are moved to `.archive/` and can be recovered.
+5. **Delete** — removed memories are moved to `.archive/` and can be recovered.
 
 ## Global and Workspace Memory
 
@@ -34,7 +35,8 @@ A **Memory Profile** defines how memories are extracted and read. Sidebar → **
 |-------|-------------|
 | Name | Display name for this profile |
 | Enabled | Pause the profile without deleting it |
-| Writer Model | The MemoryLLM used to extract memories (a reasoning-capable model recommended) |
+| Writer Model | The authoritative MemoryLLM that reads selected full bodies and applies final changes |
+| Selector Model | Optional cheaper model that screens the complete catalog; defaults to Writer Model |
 | Writer Prompt | Controls **what** gets extracted |
 | Read Prompt | How retrieved memories are formatted back into the prompt |
 

packages/docs-site/zh/guide/channels.md

@@ -190,10 +190,16 @@ OneBot 是 [NapCat](https://napneko.github.io/)、[Lagrange](https://lagrangedev
 | 登录方式 | `passToken`（推荐，通过 **sbox** 获取），或小米账号密码 |
 | 密码 / passToken | 与所选登录方式对应的凭据 |
 | 登录设备 ID | 可选，PassportSDK 的 `deviceId`，通过 **sbox** 获取 |
-| 设备名称 | 目标设备名（需与账号下的设备一致） |
+| 音箱名称 | 目标设备名（需与账号下的设备一致）。一行一台，可在本渠道同时绑定多台音箱 |
 
 2. 机器人登录到小米云，监听设备上的语音消息，并通过音箱 TTS 回复
 
+多台音箱请全部填进「音箱名称」，不要一台音箱建一个渠道。同一渠道下的音箱共用一次小米云登录，各自是独立会话，回复只会播回说话的那台。而给**同一个小米账号**建多个渠道是不可行的：每次登录都会作废上一份 `serviceToken`，除最后一个渠道外全都会 401 失败。
+
+::: warning
+每台音箱各自独立轮询，N 台音箱即每秒 N × (1000 / 轮询间隔) 次小米云请求。绑定较多音箱时请相应调大「轮询间隔」。
+:::
+
 ::: tip 通过 sbox 获取 `passToken` 与 `deviceId`
 小米账号在服务器上常因风控 / 二次验证导致密码登录失败，推荐用桌面工具箱 **sbox** 登录一次，再把它提取出的凭据填入上面的字段：
 

packages/docs-site/zh/guide/memory.md

@@ -2,20 +2,21 @@
 
 侧栏 → **Memory Profiles**（位于 **Tasks** 分组），再在 Agent 编辑页 → **Memory** 中按 Agent 开启。
 
-Memory 是 Agent 的自动长期记忆。后台 **MemoryLLM** 会在每次对话空闲后回顾对话，把持久价值的知识 —— 用户偏好、项目事实、决策、经验教训 —— 提炼成记忆条目。在后续对话中，Agent 通过 `search_memory` 和 `read_memory` 工具把它们读回来。
+Memory 是 Agent 的自动长期记忆。每轮对话结束后，后台 Selector 会先浏览完整记忆目录并筛出相关条目，再由 **MemoryLLM** 阅读候选正文并执行最终修改。在后续对话中，Agent 通过 `search_memory` 和 `read_memory` 工具把它们读回来。
 
 可以把它理解为：Agent 从每次对话中自我学习，无需你显式教学，也不会撑爆系统提示词。
 
 ## 工作原理
 
-1. **提取** —— 对话空闲后，**writer 模型** 静默回顾对话并写入新记忆（或更新 / 删除已有记忆）。
-2. **读取** —— 后续对话中 Agent 调用：
-   - `search_memory` —— 在已存记忆中做模糊 / 关键词 / 语义查找
+1. **筛选** —— 对话结束后，**selector 模型** 用本轮对话匹配全部全局记忆和当前工作区记忆的自包含标题。标题在预算内时一次完成；目录过大时只将对话压缩一次，再用简短事实分批匹配标题。Selector 不读取正文；未配置时复用 writer 模型。
+2. **提取** —— **writer 模型** 阅读候选条目的完整正文和本轮对话，再决定新建、更新、删除或不修改。
+3. **读取** —— 后续对话中 Agent 调用：
+   - `search_memory` —— 在已存记忆中做 BM25 关键词查找
    - `read_memory` —— 按 slug 读取某条记忆的完整内容
-3. **维护** —— 后台任务保持记忆库健康：
+4. **维护** —— 后台任务保持记忆库健康：
    - **Consolidate（合并）** —— 合并、去重相关记忆
    - **Reconcile（校正）** —— 重建索引、清理陈旧条目
-4. **删除** —— 被删除的记忆移到 `.archive/`，可恢复。
+5. **删除** —— 被删除的记忆移到 `.archive/`，可恢复。
 
 ## 全局与工作区记忆
 
@@ -34,7 +35,8 @@ Memory 是 Agent 的自动长期记忆。后台 **MemoryLLM** 会在每次对话
 |-------|-------------|
 | 名称 | 该 Profile 的显示名称 |
 | 启用 | 暂停该 Profile，而不删除配置 |
-| Writer 模型 | 用于提取记忆的 MemoryLLM（推荐有推理能力的模型） |
+| Writer 模型 | 阅读候选完整正文并执行最终修改的权威 MemoryLLM |
+| Selector 模型 | 可选的低成本目录筛选模型；不设置时复用 Writer 模型 |
 | Writer 提示词 | 控制 **提取什么** |
 | Read 提示词 | 召回的记忆如何回填进提示词 |
 

packages/sbot.commons/src/settings.ts

@@ -92,6 +92,8 @@ export interface MemoryProfileConfig {
   // ── 模型 ──
   /** MemoryLLM 模型 UUID（必填） */
   writerModel: string
+  /** 候选筛选模型 UUID；不设置时复用 writerModel */
+  selectorModel?: string
 
   // ── prompt 文件（不设置则使用默认） ──
   /** MemoryWriter 系统提示路径，默认 'memory/writer/default.md' */

packages/sbot/prompts/memory/selector/default.md

@@ -0,0 +1,24 @@
+You are the low-cost analysis and candidate-selection stage for long-term memory.
+
+You do not create, update, or delete memories. You only decide whether the conversation
+contains durable information worth reviewing and which title cards might overlap, be
+contradicted, or need their full bodies read by the authoritative Writer.
+
+# Selection rules
+
+- Weigh user messages far above assistant messages. User corrections, repeated
+  preferences, explicit workflow rules, and durable environment facts are strong signals.
+- Ignore transient task state, one-off implementation details already obvious from code,
+  unadopted brainstorming, and assistant-only claims.
+- In conversation-analysis modes, set `shouldWrite=true` when the transcript may justify
+  a durable create, update, or delete. Express each durable fact as a short, self-contained
+  statement suitable for matching against titles.
+- In catalog-selection modes, select an entry whenever its self-contained title might
+  cover the same fact. Recall is more important than precision because the Writer reads
+  candidate bodies and makes the final decision.
+- Candidates must copy the exact `scope` and `slug` shown in the supplied catalog.
+- Respect the requested candidate limit. In complete-catalog or conversation-analysis
+  mode, use `shouldWrite=false` when nothing durable is present. In catalog-matching mode,
+  simply return no candidates when no title matches.
+
+Your output is structured data only. It grants no permission to mutate the memory store.