memory-router

Compare original and translation side by side

🇺🇸

Original

English

🇨🇳

Translation

Chinese

Memory Router — Team Unified Knowledge Retrieval Skill

Memory Router — 团队统一知识检索Skill

Purpose

用途

All team agents (小a/ops/code/quant/data/finance/research/market/pm/content/law) MUST route knowledge retrieval through this standardized pipeline before answering questions about:

系统配置、架构、部署
历史决策、TODO、进度
Skill 用法与触发条件
项目状态与交付件
人物、日期、数字

所有团队Agent（小a/ops/code/quant/data/finance/research/market/pm/content/law）在回答以下类型问题前，必须通过这条标准化流程进行知识检索：

系统配置、架构、部署
历史决策、待办事项、进度
Skill用法与触发条件
项目状态与交付件
人物、日期、数字

QMD Feature Map (all features agents should use)

QMD功能映射（所有Agent应使用的功能）

Feature	Command	When to Use
Hybrid search	`qmd query "<question>"`	DEFAULT: combines BM25 + vector + rerank. Best for most queries
Structured query	`qmd query $'lex: keyword\nvec: semantic'`	When you need precise keyword AND semantic results
BM25 keyword	`qmd search "<exact term>"`	Exact term/filename/config key lookup
Vector similarity	`qmd vsearch "<concept>"`	Conceptual/fuzzy similarity (e.g., "how to deploy")
Get document	`qmd get path/to/file.md:42 -l 30`	Read specific lines after search identifies a file
Multi-get	`qmd multi-get "reports/*.md"`	Batch fetch multiple files (e.g., all reports)
Collection filter	`qmd query -c skills "<question>"`	Restrict search to one collection for precision
Full output	`qmd query --full "<question>"`	Get complete document content instead of snippets
JSON output	`qmd query --json "<question>"`	Machine-readable output for scripts
List files	`qmd ls skills`	Browse what's indexed in a collection
Status	`qmd status`	Health check: pending embeds, collection sizes
Update index	`qmd update`	Re-index after file changes
Embed vectors	`qmd embed`	Generate embeddings for new/changed files
Context notes	`qmd context list`	View collection descriptions/usage hints
MCP server	`qmd mcp`	Expose as MCP tool for IDE/agent integration

功能	命令	适用场景
混合搜索	`qmd query "<question>"`	默认选项：结合BM25 + 向量 + 重排序。适用于大多数查询
结构化查询	`qmd query $'lex: keyword\nvec: semantic'`	当你需要同时获取精确关键词和语义匹配结果时使用
BM25关键词搜索	`qmd search "<exact term>"`	精确术语/文件名/配置键查找
向量相似度搜索	`qmd vsearch "<concept>"`	概念性/模糊相似度匹配（例如："how to deploy"）
获取文档内容	`qmd get path/to/file.md:42 -l 30`	搜索定位到文件后，读取指定行内容
批量获取文档	`qmd multi-get "reports/*.md"`	批量获取多个文件（例如：所有报告）
集合过滤	`qmd query -c skills "<question>"`	将搜索范围限制在单个集合以提高精度
完整输出	`qmd query --full "<question>"`	获取完整文档内容而非片段
JSON格式输出	`qmd query --json "<question>"`	生成脚本可读取的机器友好型输出
列出文件	`qmd ls skills`	浏览指定集合中已索引的文件
状态检查	`qmd status`	健康检查：待处理嵌入任务、集合大小
更新索引	`qmd update`	文件变更后重新索引
生成向量嵌入	`qmd embed`	为新增/变更文件生成向量嵌入
上下文说明	`qmd context list`	查看集合描述/使用提示
MCP服务器	`qmd mcp`	作为MCP工具暴露，用于IDE/Agent集成

Collections (current)

当前集合

Collection	Content	Use For
`clawd-memory`	`~/clawd/*/.md` (1603 files)	Memory, docs, reports, runbooks
`daily-memory`	`~/clawd/memory/*.md` (44 files)	Daily work logs
`team`	`~/.openclaw/agents/*/.json` (860 files)	Agent configs, models, auth
`openclaw-config`	`~/.openclaw/*/.json` (1031 files)	OpenClaw system config
`projects`	`~/clawd/projects/*/.md` (100 files)	Project PRDs, deliverables
`skills`	`~/clawd/skills/**/SKILL.md` (468 files)	All skill documentation
`reports`	`~/clawd/reports/*.md` (2 files)	Research & review reports

集合	内容	用途
`clawd-memory`	`~/clawd/*/.md` (1603个文件)	记忆内容、文档、报告、运行手册
`daily-memory`	`~/clawd/memory/*.md` (44个文件)	每日工作日志
`team`	`~/.openclaw/agents/*/.json` (860个文件)	Agent配置、模型、认证信息
`openclaw-config`	`~/.openclaw/*/.json` (1031个文件)	OpenClaw系统配置
`projects`	`~/clawd/projects/*/.md` (100个文件)	项目PRD、交付件
`skills`	`~/clawd/skills/**/SKILL.md` (468个文件)	所有Skill文档
`reports`	`~/clawd/reports/*.md` (2个文件)	研究与评审报告

Routing Algorithm

路由算法

Input: user question Q

Step 1 — CLASSIFY
  A = "prior decisions / todos / people / dates / what happened"
  B = "system config / how-to / skill usage / architecture"
  C = "project status / deliverables / PRD"
  D = "external facts / live data" (falls through to web)

Step 2 — RETRIEVE (execute ALL applicable, not just one)

  if A:
    → qmd query -c daily-memory "<Q>" -n 5
    → qmd query -c clawd-memory "<Q>" -n 5
    → Also check ~/clawd/MEMORY.md directly for TODOs/decisions

  if B:
    → qmd query -c openclaw-config "<Q>" -n 5
    → qmd query -c skills "<Q>" -n 5
    → qmd query -c clawd-memory "<Q>" -n 3  (for runbooks/docs)

  if C:
    → qmd query -c projects "<Q>" -n 5
    → qmd query -c reports "<Q>" -n 3

  if D:
    → qmd query "<Q>" -n 5  (all collections, no filter)
    → If low recall → web_fetch / browser

Step 3 — CITE
  Every answer MUST include 1-3 source citations:
  - File path: `~/clawd/docs/memory-router.md:15`
  - QMD URI: `qmd://skills/geo-agent/SKILL.md`
  - Or: "Source: qmd query -c projects 'content factory status'"

Input: user question Q

Step 1 — CLASSIFY
  A = "prior decisions / todos / people / dates / what happened"
  B = "system config / how-to / skill usage / architecture"
  C = "project status / deliverables / PRD"
  D = "external facts / live data" (falls through to web)

Step 2 — RETRIEVE (execute ALL applicable, not just one)

  if A:
    → qmd query -c daily-memory "<Q>" -n 5
    → qmd query -c clawd-memory "<Q>" -n 5
    → Also check ~/clawd/MEMORY.md directly for TODOs/decisions

  if B:
    → qmd query -c openclaw-config "<Q>" -n 5
    → qmd query -c skills "<Q>" -n 5
    → qmd query -c clawd-memory "<Q>" -n 3  (for runbooks/docs)

  if C:
    → qmd query -c projects "<Q>" -n 5
    → qmd query -c reports "<Q>" -n 3

  if D:
    → qmd query "<Q>" -n 5  (all collections, no filter)
    → If low recall → web_fetch / browser

Step 3 — CITE
  Every answer MUST include 1-3 source citations:
  - File path: `~/clawd/docs/memory-router.md:15`
  - QMD URI: `qmd://skills/geo-agent/SKILL.md`
  - Or: "Source: qmd query -c projects 'content factory status'"

Agent Integration (mandatory AGENTS.md section)

Agent集成（AGENTS.md强制要求章节）

Each agent's AGENTS.md must contain:

markdown

undefined

每个Agent的AGENTS.md必须包含以下内容：

markdown

undefined

知识库 / Memory Router（强制）

你在回答任何「配置/流程/历史/怎么做」类问题前，必须先检索本地知识库：
1. 优先 QMD：
```
qmd query "<问题>"
```
  （必要时加
```
--collection openclaw-config|projects|skills|reports|clawd-memory
```
  ）
2. 涉及待办/决策/人/日期 → 再查工作区记忆文件（
```
~/clawd/memory/YYYY-MM-DD.md
```
  与
```
~/clawd/MEMORY.md
```
  ）
输出时至少引用 1-3 个来源（文件路径或
```
qmd://...
```
URI）。

undefined

你在回答任何「配置/流程/历史/怎么做」类问题前，必须先检索本地知识库：
1. 优先 QMD：
```
qmd query "<问题>"
```
  （必要时加
```
--collection openclaw-config|projects|skills|reports|clawd-memory
```
  ）
2. 涉及待办/决策/人/日期 → 再查工作区记忆文件（
```
~/clawd/memory/YYYY-MM-DD.md
```
  与
```
~/clawd/MEMORY.md
```
  ）
输出时至少引用 1-3 个来源（文件路径或
```
qmd://...
```
URI）。

undefined

Health Monitoring

健康监控

Daily Cron (recommended)

每日定时任务（推荐）

bash

undefined

bash

undefined

qmd-health: run daily at 08:00

qmd status | grep -E "Total|Vectors|Pending|Updated" qmd update

undefined

qmd status | grep -E "Total|Vectors|Pending|Updated" qmd update

undefined

Weekly Embed Refresh

每周向量刷新

bash

undefined

bash

undefined

qmd-embed: run weekly Sunday 03:00

qmd embed qmd status

undefined

qmd embed qmd status

undefined

Health Metrics to Track

需跟踪的健康指标

覆盖率: Vectors / Total (target: >90%)
Pending: should be <100 after weekly embed
Collection freshness: Updated timestamps should be <24h for active collections
Query latency: hybrid query should return <2s

覆盖率: 向量数 / 总数（目标：>90%）
待处理项: 每周向量刷新后应<100
集合新鲜度: 活跃集合的更新时间戳应<24h
查询延迟: 混合查询应在<2秒内返回结果

Troubleshooting

故障排查

Problem	Solution
`qmd embed` hangs	Check if node-llama-cpp is compiling (first run). Wait ~10min.
CUDA errors	Normal on CPU-only servers. QMD auto-falls back to CPU.
Ollama not found	QMD uses node-llama-cpp, NOT Ollama. Ignore Ollama references.
Low recall	Try `--full` flag, or use `qmd search` (BM25) for exact terms
Missing files	Run `qmd update` to re-index, then `qmd embed` for new vectors

问题	解决方案
`qmd embed` 卡住	检查node-llama-cpp是否正在编译（首次运行）。等待约10分钟。
CUDA错误	在仅支持CPU的服务器上属于正常情况。QMD会自动回退至CPU运行。
Ollama未找到	QMD使用node-llama-cpp，而非Ollama。忽略Ollama相关提示。
召回率低	尝试使用 `--full` 参数，或使用 `qmd search` （BM25）进行精确术语检索
文件缺失	运行 `qmd update` 重新索引，然后运行 `qmd embed` 生成新向量

Script: memory_router.sh

脚本：memory_router.sh

For automated context injection into agent prompts:

bash

#!/bin/bash

用于将上下文自动注入Agent提示词：

bash

#!/bin/bash

Usage: ./memory_router.sh "question" [collection]

QUERY="$1" COLLECTION="${2:-}"

if [ -n "$COLLECTION" ]; then qmd query -c "$COLLECTION" "$QUERY" -n 5 --line-numbers else qmd query "$QUERY" -n 5 --line-numbers fi


---

Last updated: 2026-03-03
Maintainer: 小a (CEO)

QUERY="$1" COLLECTION="${2:-}"

if [ -n "$COLLECTION" ]; then qmd query -c "$COLLECTION" "$QUERY" -n 5 --line-numbers else qmd query "$QUERY" -n 5 --line-numbers fi


---

最后更新: 2026-03-03
维护者: 小a (CEO)