feat: recallMode with adaptive intent routing and tiered injection

CHANGELOG.md

@@ -1,5 +1,14 @@
 # Changelog
 
+## 1.1.0-beta.5
+
+- Refactor: build reset/new reflection handoff note in `runMemoryReflection`.
+- Refactor: `<open-loops>` now comes from the fresh reflection run, while `<derived-focus>` comes from historical scored itemized derived rows.
+- Breaking: stop writing and stop reading legacy combined reflection rows (`type=memory-reflection`).
+- Docs: refresh README / README_CN for the new handoff-note behavior and remove old legacy combined guidance.
+
+---
+
 ## 1.1.0
 
 - Feat: add integrated self-improvement governance flow (`agent:bootstrap`, `command:new/reset`, governance tools, and `.learnings` file bootstrap).

README.md

@@ -160,6 +160,10 @@ Filters out low-quality content at both auto-capture and tool-store stages:
 - Hooks: `agent:bootstrap`, `command:new`, `command:reset`.
 - `agent:bootstrap`: injects `SELF_IMPROVEMENT_REMINDER.md` into bootstrap context.
 - `command:new` / `command:reset`: appends a short `/note self-improvement ...` reminder before reset.
+  - Under `sessionStrategy="memoryReflection"`, the final reset/new note is assembled in `runMemoryReflection`.
+  - In `memoryReflection.injectMode="inheritance+derived"` mode, the note can include:
+    - `<open-loops>` from the **fresh** reflection text of the current `/new` or `/reset`.
+    - `<derived-focus>` from **historical** LanceDB derived rows after dedupe+decay scoring, filtered to final score `> 0.3`.
 - File ensure/create path: ensures `.learnings/LEARNINGS.md` and `.learnings/ERRORS.md` exist.
 - This flow is separate from `memoryReflection`: seeing self-improvement notes or `.learnings/*` activity does not by itself mean reflection storage is enabled.
 - Append paths are intentionally distinct:
@@ -183,6 +187,7 @@ Filters out low-quality content at both auto-capture and tool-store stages:
   - Requires `sessionStrategy="memoryReflection"`.
   - Triggers on `command:new` / `command:reset`.
   - Skips generation when session context is incomplete (for example missing config, session file, or readable conversation content).
+  - Edge case: issuing `/new` immediately after a previous `/new` can land in a fresh empty session without a readable prior `sessionFile`; in that case reflection skips and logs `missing session file after recovery`. This is expected behavior.
 - Runner chain:
   - First try embedded runner (`runEmbeddedPiAgent`).
   - If the embedded path fails, fall back to `openclaw agent --local --json`.
@@ -201,7 +206,6 @@ Filters out low-quality content at both auto-capture and tool-store stages:
   - Writes one event row (`type=memory-reflection-event`) plus item rows (`type=memory-reflection-item`) for each `Invariants` / `Derived` bullet.
   - Event rows keep lightweight provenance/audit metadata only (`eventId`, `sessionKey`, `usedFallback`, `errorSignals`, source path).
   - Item rows carry per-item decay metadata (`decayModel`, `decayMidpointDays`, `decayK`, `baseWeight`, `quality`) plus ordinal/group metadata.
-  - Compatibility mode: `memoryReflection.writeLegacyCombined=true` (default) also writes legacy combined rows (`type=memory-reflection`) during migration.
   - Reflection rows display as `reflection:<scope>`.
 - Reflection-derived durable memory mapping:
   - Available memory categories in the plugin are `preference`, `fact`, `decision`, `entity`, `reflection`, `other`.
@@ -229,6 +233,12 @@ Filters out low-quality content at both auto-capture and tool-store stages:
   - If the configured agent id is not present in `cfg.agents.list`, the plugin warns and falls back to the runtime agent id.
 - Error loop:
   - `after_tool_call` captures and deduplicates tool error signatures for reminder/reflection context.
+- Injection placement by hook (`memoryReflection.injectMode`):
+  - `before_agent_start`: injects `<inherited-rules>` via Reflection-Recall.
+  - `memoryReflection.recall.mode="fixed"` (default): compatibility path; fixed inheritance remains active even when generic Auto-Recall is disabled.
+  - `memoryReflection.recall.mode="dynamic"`: prompt-gated dynamic Reflection-Recall with independent top-k/session suppression budget from generic Auto-Recall.
+  - `command:new` / `command:reset`: `runMemoryReflection` builds the self-improvement note (`<open-loops>` from fresh reflection; `<derived-focus>` from historical scored rows when mode is `inheritance+derived`).
+  - `before_prompt_build`: injects `<error-detected>` only (no `<derived-focus>`).
 
 ### 10. Markdown Mirror (`mdMirror`)
 
@@ -274,7 +284,11 @@ When embedding calls fail, the plugin provides **actionable error messages** ins
 
 - **Auto-Capture** (`agent_end` hook): Extracts preference/fact/decision/entity from conversations, deduplicates, stores up to 3 per turn
   - Skips memory-management prompts (e.g. delete/forget/cleanup memory entries) to reduce noise
-- **Auto-Recall** (`before_agent_start` hook): Injects `<relevant-memories>` context (up to 3 entries)
+- **Auto-Recall** (`before_agent_start` hook): Injects `<relevant-memories>` context
+  - Default top-k: `autoRecallTopK=3`
+  - Default category allowlist: `preference`, `fact`, `decision`, `entity`, `other`
+  - `autoRecallExcludeReflection=true` by default, so `<relevant-memories>` stays separate from `<inherited-rules>`
+  - Supports age window (`autoRecallMaxAgeDays`) and recent-per-key cap (`autoRecallMaxEntriesPerKey`)
 
 ### Prevent memories from showing up in replies
 
@@ -309,7 +323,7 @@ Add a line to your agent system prompt, e.g.:
 
 ## Installation
 
-> **🧪 Beta available: v1.1.0-beta.3**
+> **🧪 Beta available: v1.1.0-beta.5**
 >
 > A beta release is available with major new features: **Self-Improvement governance**, **memoryReflection session strategy**, **Markdown Mirror**, and improved embedding error diagnostics. The stable `latest` remains at v1.0.32.
 >
@@ -321,7 +335,7 @@ Add a line to your agent system prompt, e.g.:
 > npm install memory-lancedb-pro
 > ```
 >
-> See [Release Notes](https://github.com/win4r/memory-lancedb-pro/releases/tag/v1.1.0-beta.3) for details. Feedback welcome via [GitHub Issues](https://github.com/win4r/memory-lancedb-pro/issues).
+> See [Release Notes](https://github.com/win4r/memory-lancedb-pro/releases/tag/v1.1.0-beta.5) for details. Feedback welcome via [GitHub Issues](https://github.com/win4r/memory-lancedb-pro/issues).
 
 ### AI-safe install notes (anti-hallucination)
 
@@ -470,6 +484,11 @@ openclaw config get plugins.slots.memory
   "autoCapture": true,
   "autoRecall": false,
   "autoRecallMinLength": 8,
+  "autoRecallTopK": 3,
+  "autoRecallCategories": ["preference", "fact", "decision", "entity", "other"],
+  "autoRecallExcludeReflection": true,
+  "autoRecallMaxAgeDays": 30,
+  "autoRecallMaxEntriesPerKey": 10,
   "retrieval": {
     "mode": "hybrid",
     "vectorWeight": 0.7,
@@ -510,15 +529,24 @@ openclaw config get plugins.slots.memory
   },
   "memoryReflection": {
     "storeToLanceDB": true,
-    "writeLegacyCombined": true,
     "injectMode": "inheritance+derived",
     "agentId": "memory-distiller",
     "messageCount": 120,
     "maxInputChars": 24000,
     "timeoutMs": 20000,
     "thinkLevel": "medium",
     "errorReminderMaxEntries": 3,
-    "dedupeErrorSignals": true
+    "dedupeErrorSignals": true,
+    "recall": {
+      "mode": "fixed",
+      "topK": 6,
+      "includeKinds": ["invariant"],
+      "maxAgeDays": 45,
+      "maxEntriesPerKey": 10,
+      "minRepeated": 2,
+      "minScore": 0.18,
+      "minPromptLength": 8
+    }
   },
   "mdMirror": {
     "enabled": false,
@@ -547,6 +575,7 @@ A practical starting point for Chinese chat workloads:
   "autoCapture": true,
   "autoRecall": true,
   "autoRecallMinLength": 8,
+  "autoRecallExcludeReflection": true,
   "retrieval": {
     "candidatePoolSize": 20,
     "minScore": 0.45,

README_CN.md

@@ -183,6 +183,7 @@ Query → BM25 FTS ─────┘
   - 需要 `sessionStrategy="memoryReflection"`。
   - 触发于 `command:new` / `command:reset`。
   - 若会话上下文不完整（例如缺少配置、session 文件、或可读对话内容），则跳过生成。
+  - 边界场景：在刚 `/new` 后立即再次 `/new`，可能进入一个没有可读上一轮 `sessionFile` 的新空会话；此时 reflection 会跳过并记录 `missing session file after recovery`。这属于预期行为。
 - 执行链：
   - 先尝试 embedded runner（`runEmbeddedPiAgent`）。
   - 若 embedded 路径失败，则回退到 `openclaw agent --local --json`。
@@ -201,7 +202,7 @@ Query → BM25 FTS ─────┘
   - 每次反思会写入 1 条 event row（`type=memory-reflection-event`）以及多条 item row（`type=memory-reflection-item`，对应每个 `Invariants` / `Derived` bullet）。
   - event row 仅保留轻量 provenance / audit 元数据（`eventId`、`sessionKey`、`usedFallback`、`errorSignals`、source path）。
   - item row 携带逐条衰减元数据（`decayModel`、`decayMidpointDays`、`decayK`、`baseWeight`、`quality`）以及 `ordinal/groupSize`。
-  - 兼容模式：`memoryReflection.writeLegacyCombined=true`（默认）时，迁移期仍会额外写入旧版 combined row（`type=memory-reflection`）。
+  - 仅使用 itemized reflection rows；旧版 combined row（`type=memory-reflection`）已不再写入，也不再参与读取。
   - reflection 行展示标签为 `reflection:<scope>`。
 - Reflection 派生 durable memory 映射：
   - 插件支持的 memory categories 包括 `preference`、`fact`、`decision`、`entity`、`reflection`、`other`。
@@ -229,6 +230,12 @@ Query → BM25 FTS ─────┘
   - 若配置的 agent id 不在 `cfg.agents.list` 中，插件会告警并回退到 runtime agent id。
 - 错误闭环：
   - `after_tool_call` 捕获并去重工具错误签名，用于提醒 / reflection 上下文。
+- Hook 注入位置（`memoryReflection.injectMode`）：
+  - `before_agent_start`：通过 Reflection-Recall 注入 `<inherited-rules>`。
+  - `memoryReflection.recall.mode="fixed"`（默认）：兼容路径；即使关闭 generic Auto-Recall，固定继承仍会注入。
+  - `memoryReflection.recall.mode="dynamic"`：按 prompt 动态检索反思规则，且与 generic Auto-Recall 使用独立 top-k / session 去重预算。
+  - `command:new` / `command:reset`：`runMemoryReflection` 生成 self-improvement note（`<open-loops>` 来自本次新反思；`<derived-focus>` 来自历史打分行，模式为 `inheritance+derived` 时启用）。
+  - `before_prompt_build`：仅注入 `<error-detected>`（不会注入 `<derived-focus>`）。
 
 ### 10. Markdown 镜像（`mdMirror`）
 
@@ -274,7 +281,11 @@ Query → BM25 FTS ─────┘
 
 - **Auto-Capture**（`agent_end` hook）: 从对话中提取 preference/fact/decision/entity，去重后存储（每次最多 3 条）
   - 触发词支持 **简体中文 + 繁體中文**（例如：记住/記住、偏好/喜好/喜歡、决定/決定 等）
-- **Auto-Recall**（`before_agent_start` hook）: 注入 `<relevant-memories>` 上下文（最多 3 条）
+- **Auto-Recall**（`before_agent_start` hook）: 注入 `<relevant-memories>` 上下文
+  - 默认 top-k：`autoRecallTopK=3`
+  - 默认类别白名单：`preference`、`fact`、`decision`、`entity`、`other`
+  - 默认 `autoRecallExcludeReflection=true`，让 `<relevant-memories>` 与 `<inherited-rules>` 分离
+  - 支持时间窗（`autoRecallMaxAgeDays`）和按归一化 key 的最近 N 条限制（`autoRecallMaxEntriesPerKey`）
 
 ### 不想在对话中“显示长期记忆”？
 
@@ -309,7 +320,7 @@ Query → BM25 FTS ─────┘
 
 ## 安装
 
-> **🧪 Beta 版本可用：v1.1.0-beta.3**
+> **🧪 Beta 版本可用：v1.1.0-beta.5**
 >
 > Beta 版包含多项重大新特性：**Self-Improvement 治理流**、**memoryReflection 会话策略**、**Markdown 镜像双写**、以及改进的 Embedding 错误诊断。稳定版 `latest` 仍为 v1.0.32。
 >
@@ -321,7 +332,7 @@ Query → BM25 FTS ─────┘
 > npm install memory-lancedb-pro
 > ```
 >
-> 详见 [Release Notes](https://github.com/win4r/memory-lancedb-pro/releases/tag/v1.1.0-beta.3)。欢迎通过 [GitHub Issues](https://github.com/win4r/memory-lancedb-pro/issues) 反馈问题。
+> 详见 [Release Notes](https://github.com/win4r/memory-lancedb-pro/releases/tag/v1.1.0-beta.5)。欢迎通过 [GitHub Issues](https://github.com/win4r/memory-lancedb-pro/issues) 反馈问题。
 
 ### AI 安装指引（防幻觉版）
 
@@ -470,6 +481,11 @@ openclaw config get plugins.slots.memory
   "autoCapture": true,
   "autoRecall": false,
   "autoRecallMinLength": 8,
+  "autoRecallTopK": 3,
+  "autoRecallCategories": ["preference", "fact", "decision", "entity", "other"],
+  "autoRecallExcludeReflection": true,
+  "autoRecallMaxAgeDays": 30,
+  "autoRecallMaxEntriesPerKey": 10,
   "retrieval": {
     "mode": "hybrid",
     "vectorWeight": 0.7,
@@ -510,15 +526,24 @@ openclaw config get plugins.slots.memory
   },
   "memoryReflection": {
     "storeToLanceDB": true,
-    "writeLegacyCombined": true,
     "injectMode": "inheritance+derived",
     "agentId": "memory-distiller",
     "messageCount": 120,
     "maxInputChars": 24000,
     "timeoutMs": 20000,
     "thinkLevel": "medium",
     "errorReminderMaxEntries": 3,
-    "dedupeErrorSignals": true
+    "dedupeErrorSignals": true,
+    "recall": {
+      "mode": "fixed",
+      "topK": 6,
+      "includeKinds": ["invariant"],
+      "maxAgeDays": 45,
+      "maxEntriesPerKey": 10,
+      "minRepeated": 2,
+      "minScore": 0.18,
+      "minPromptLength": 8
+    }
   },
   "mdMirror": {
     "enabled": false,
@@ -547,6 +572,7 @@ openclaw config get plugins.slots.memory
   "autoCapture": true,
   "autoRecall": true,
   "autoRecallMinLength": 8,
+  "autoRecallExcludeReflection": true,
   "retrieval": {
     "candidatePoolSize": 20,
     "minScore": 0.45,