最近不少用户在讨论使用MemOS这类付费插件来降低OpenClaw的Token消耗。但深入研究其源码后我发现，它的核心逻辑其实并不复杂，完全可以用一套纯本地的方案来实现，甚至效果更好。
本文将手把手教你搭建一个零费用、零云端依赖的记忆增强系统，全程在Cursor中操作，无需编码基础也能轻松搞定。
这个方案能帮你解决什么？
OpenClaw原生的记忆系统存在几个明显的痛点：
- 全局记忆膨胀：
MEMORY.md 文件会越来越大，每次对话都要加载全文，导致无效的Token消耗。
- 记忆依赖模型自觉：Agent经常“忘记”主动记录关键信息。
- 记忆召回不精准：要么把所有记忆都塞进上下文，要么完全找不到相关记忆。
我们的本地增强方案致力于解决这些问题：
- 自动记录：每次对话结束后，自动将内容存入本地数据库，不依赖模型主动记忆。
- 按需召回：对话开始前，自动搜索相关的近期对话、摘要和用户偏好，只注入必要的部分。
- 偏好学习：通过夜间任务自动分析你的历史对话，提取并学习你的使用习惯和偏好。
- 完全本地化：基于 SQLite 数据库和本地的 Ollama 模型，零云端依赖，零费用。
工作原理
整个系统的工作流清晰分为两条路径：
1. 实时路径 (每次对话触发)
用户发送消息
→ 触发 `before_agent_start` 钩子
→ 通过 QMD 搜索长期记忆（workspace/memory/ 下的文件）
→ 从 SQLite 搜索近期相关对话与摘要
→ 从 SQLite 读取用户偏好信息
→ 将所有信息拼装成 `prependContext`，注入到提示词前部
→ [AI Agent](https://yunpan.plus/f/29-1) 基于已有记忆上下文进行回复（无需再读取庞大的 memory 文件）
→ 触发 `agent_end` 钩子
→ 将本轮对话存储到 SQLite（毫秒级 INSERT 操作）
2. 夜间路径 (每日自动执行)
定时任务（Cron Job）触发
→ 读取当天所有未处理的对话记录
→ 发送给本地运行的 Ollama 模型进行分析
→ 提取用户偏好 → 写入 preferences 表
→ 生成对话摘要 → 写入 summaries 表
→ 标记这些对话为已处理
前置条件
开始之前，请确保你的环境已满足以下要求：
- 已安装 OpenClaw（版本 2026.2.13 或更高）。
- 已安装并配置好 QMD 插件。
- 已安装 Ollama（用于夜间偏好提取），安装命令：
brew install ollama。
- 已在 Ollama 中拉取一个小模型，例如：
ollama pull qwen3:8b。
- 准备好 Cursor IDE（用于跟随教程操作）。
第一步：创建插件目录
打开 Cursor 的终端，运行以下命令创建必要的目录结构：
mkdir -p ~/.openclaw/extensions/local-memory-recall/lib
mkdir -p ~/.openclaw/extensions/local-memory-recall/scripts
mkdir -p ~/.openclaw/extensions/local-memory-recall/data
第二步：创建插件描述文件
在 Cursor 中新建文件，路径为：~/.openclaw/extensions/local-memory-recall/openclaw.plugin.json
将以下 JSON 配置内容粘贴进去：
{
"id": "local-memory-recall",
"name": "Local Memory Recall",
"description": "Auto-capture conversations and recall relevant memories via local SQLite + QMD. Zero cloud dependencies, zero cost.",
"version": "1.0.0",
"kind": "lifecycle",
"main": "./index.js",
"configSchema": {
"type": "object",
"properties": {
"agentIds": {
"type": "array",
"items": { "type": "string" },
"description": "Only activate for these agent IDs. Empty = all agents."
},
"recallEnabled": { "type": "boolean", "default": true },
"addEnabled": { "type": "boolean", "default": true },
"memoryLimit": { "type": "integer", "default": 6 },
"conversationLimit": { "type": "integer", "default": 5 },
"preferenceLimit": { "type": "integer", "default": 6 },
"maxDaysBack": { "type": "integer", "default": 30 },
"captureStrategy": {
"type": "string",
"enum": ["last_turn", "full_session"],
"default": "last_turn"
},
"maxMessageChars": { "type": "integer", "default": 20000 },
"dbPath": { "type": "string" }
},
"additionalProperties": false
}
}
第三步：创建 SQLite 存储层
新建文件，路径为：~/.openclaw/extensions/local-memory-recall/lib/local-store.js
这是核心的数据库操作层，负责所有数据的存储和检索，请完整粘贴以下代码：
import { DatabaseSync } from "node:sqlite";
import { mkdirSync, existsSync } from "node:fs";
import { dirname, join } from "node:path";
const DEFAULT_DB_DIR = join(dirname(new URL(import.meta.url).pathname), "..", "data");
export class LocalMemoryStore {
#db;
constructor(dbPath) {
const resolvedPath = dbPath || join(DEFAULT_DB_DIR, "memories.db");
const dir = dirname(resolvedPath);
if (!existsSync(dir)) mkdirSync(dir, { recursive: true });
this.#db = new DatabaseSync(resolvedPath);
this.#db.exec("PRAGMA journal_mode=WAL");
this.#db.exec("PRAGMA busy_timeout=3000");
this.#migrate();
}
#migrate() {
this.#db.exec(`
CREATE TABLE IF NOT EXISTS conversations (
id INTEGER PRIMARY KEY AUTOINCREMENT,
session_id TEXT,
agent_id TEXT NOT NULL,
role TEXT NOT NULL,
content TEXT NOT NULL,
created_at TEXT NOT NULL DEFAULT (datetime('now')),
processed INTEGER NOT NULL DEFAULT 0
);
CREATE INDEX IF NOT EXISTS idx_conv_agent_created
ON conversations(agent_id, created_at DESC);
CREATE INDEX IF NOT EXISTS idx_conv_processed
ON conversations(processed, created_at);
CREATE TABLE IF NOT EXISTS preferences (
id INTEGER PRIMARY KEY AUTOINCREMENT,
agent_id TEXT NOT NULL,
preference TEXT NOT NULL,
type TEXT NOT NULL DEFAULT 'explicit',
source_conversation_id INTEGER,
created_at TEXT NOT NULL DEFAULT (datetime('now')),
updated_at TEXT NOT NULL DEFAULT (datetime('now')),
UNIQUE(agent_id, preference)
);
CREATE INDEX IF NOT EXISTS idx_pref_agent ON preferences(agent_id);
CREATE TABLE IF NOT EXISTS summaries (
id INTEGER PRIMARY KEY AUTOINCREMENT,
conversation_group_id TEXT NOT NULL,
agent_id TEXT NOT NULL,
summary TEXT NOT NULL,
created_at TEXT NOT NULL DEFAULT (datetime('now'))
);
CREATE INDEX IF NOT EXISTS idx_sum_agent_created
ON summaries(agent_id, created_at DESC);
`);
}
addConversation({ sessionId, agentId, role, content }) {
this.#db.prepare(
`INSERT INTO conversations (session_id, agent_id, role, content) VALUES (?, ?, ?, ?)`
).run(sessionId || "", agentId, role, content);
}
upsertPreference({ agentId, preference, type, sourceConversationId }) {
this.#db.prepare(
`INSERT INTO preferences (agent_id, preference, type, source_conversation_id)
VALUES (?, ?, ?, ?)
ON CONFLICT(agent_id, preference) DO UPDATE SET
type = excluded.type, updated_at = datetime('now')`
).run(agentId, preference, type || "explicit", sourceConversationId || null);
}
addSummary({ conversationGroupId, agentId, summary }) {
this.#db.prepare(
`INSERT INTO summaries (conversation_group_id, agent_id, summary) VALUES (?, ?, ?)`
).run(conversationGroupId, agentId, summary);
}
markConversationsProcessed(ids) {
if (!ids?.length) return;
const placeholders = ids.map(() => "?").join(",");
this.#db.prepare(
`UPDATE conversations SET processed = 1 WHERE id IN (${placeholders})`
).run(...ids);
}
searchConversations({ agentId, query, limit = 5, maxDaysBack = 30 }) {
const summaries = this.#searchSummaries({ agentId, query, limit, maxDaysBack });
if (summaries.length >= limit) return summaries;
const remaining = limit - summaries.length;
const conversations = this.#searchRawConversations({ agentId, query, limit: remaining, maxDaysBack });
return [...summaries, ...conversations];
}
#searchSummaries({ agentId, query, limit, maxDaysBack }) {
const keywords = this.#extractKeywords(query);
if (!keywords.length) {
return this.#db.prepare(
`SELECT summary AS content, 'summary' AS source, created_at FROM summaries
WHERE agent_id = ? AND created_at > datetime('now', ?) ORDER BY created_at DESC LIMIT ?`
).all(agentId, `-${maxDaysBack} days`, limit);
}
const likeClause = keywords.map(() => "summary LIKE ?").join(" OR ");
const likeParams = keywords.map((k) => `%${k}%`);
return this.#db.prepare(
`SELECT summary AS content, 'summary' AS source, created_at FROM summaries
WHERE agent_id = ? AND created_at > datetime('now', ?) AND (${likeClause})
ORDER BY created_at DESC LIMIT ?`
).all(agentId, `-${maxDaysBack} days`, ...likeParams, limit);
}
#searchRawConversations({ agentId, query, limit, maxDaysBack }) {
const keywords = this.#extractKeywords(query);
if (!keywords.length) {
return this.#db.prepare(
`SELECT role, content, 'conversation' AS source, created_at FROM conversations
WHERE agent_id = ? AND created_at > datetime('now', ?) ORDER BY created_at DESC LIMIT ?`
).all(agentId, `-${maxDaysBack} days`, limit);
}
const likeClause = keywords.map(() => "content LIKE ?").join(" OR ");
const likeParams = keywords.map((k) => `%${k}%`);
return this.#db.prepare(
`SELECT role, content, 'conversation' AS source, created_at FROM conversations
WHERE agent_id = ? AND created_at > datetime('now', ?) AND (${likeClause})
ORDER BY created_at DESC LIMIT ?`
).all(agentId, `-${maxDaysBack} days`, ...likeParams, limit);
}
getPreferences({ agentId, limit = 6 }) {
return this.#db.prepare(
`SELECT preference, type, updated_at FROM preferences
WHERE agent_id = ? ORDER BY updated_at DESC LIMIT ?`
).all(agentId, limit);
}
getUnprocessedConversations({ limit = 200 } = {}) {
return this.#db.prepare(
`SELECT id, session_id, agent_id, role, content, created_at FROM conversations
WHERE processed = 0 ORDER BY created_at ASC LIMIT ?`
).all(limit);
}
cleanup() {
this.#db.exec(`DELETE FROM conversations WHERE created_at < datetime('now', '-30 days')`);
this.#db.exec(`DELETE FROM summaries WHERE created_at < datetime('now', '-90 days')`);
}
close() { this.#db.close(); }
#extractKeywords(query) {
if (!query) return [];
return query.replace(/[^\p{L}\p{N}\s]/gu, " ").split(/\s+/).filter((w) => w.length >= 2).slice(0, 8);
}
}
第四步：创建主插件文件
新建文件，路径为：~/.openclaw/extensions/local-memory-recall/index.js
这是插件的主逻辑文件，负责钩子事件的监听与处理，请完整粘贴以下代码：
import { LocalMemoryStore } from "./lib/local-store.js";
const PLUGIN_TAG = "[local-memory-recall]";
function truncate(text, maxLen) {
if (!text || !maxLen) return text || "";
return text.length > maxLen ? `${text.slice(0, maxLen)}...` : text;
}
function extractText(content) {
if (!content) return "";
if (typeof content === "string") return content;
if (Array.isArray(content)) {
return content.filter((b) => b?.type === "text").map((b) => b.text).join(" ");
}
return "";
}
function pad2(v) { return String(v).padStart(2, "0"); }
function formatTime(ts) {
const d = ts ? new Date(ts) : new Date();
return `${d.getFullYear()}-${pad2(d.getMonth() + 1)}-${pad2(d.getDate())} ${pad2(d.getHours())}:${pad2(d.getMinutes())}`;
}
function pickLastTurnMessages(messages, maxChars) {
const lastUserIdx = messages.map((m, i) => ({ m, i })).filter(({ m }) => m?.role === "user").map(({ i }) => i).pop();
if (lastUserIdx === undefined) return [];
const results = [];
for (const msg of messages.slice(lastUserIdx)) {
if (!msg?.role) continue;
const content = extractText(msg.content);
if (!content) continue;
if (msg.role === "user" || msg.role === "assistant") {
results.push({ role: msg.role, content: truncate(content, maxChars) });
}
}
return results;
}
function buildPrependContext({ qmdResults, conversations, preferences, nowText }) {
const lines = [];
lines.push("# Role", "",
"You are an intelligent assistant with long-term memory. " +
"Use the retrieved memory fragments below to provide personalized, accurate responses.", "");
lines.push("# System Context", "", `* Current Time: ${nowText}`, "");
lines.push("# Memory Data", "");
if (qmdResults?.length) {
lines.push("## Long-term Memory (from knowledge base)", "```text", "<facts>");
for (const item of qmdResults) {
const snippet = item.snippet || item.content || "";
if (snippet) lines.push(` - ${snippet.replace(/\n/g, " ").trim()}`);
}
lines.push("</facts>", "```", "");
}
if (conversations?.length) {
lines.push("## Recent Conversations", "```text", "<recent_context>");
for (const item of conversations) {
const time = item.created_at || "";
if (item.source === "summary") {
lines.push(` - [${time}] (summary) ${item.content}`);
} else {
lines.push(` - [${time}] [${item.role}] ${truncate(item.content, 500)}`);
}
}
lines.push("</recent_context>", "```", "");
}
if (preferences?.length) {
lines.push("## User Preferences", "```text", "<preferences>");
for (const pref of preferences) {
const typeLabel = pref.type === "implicit" ? "[Implicit]" : "[Explicit]";
lines.push(` - ${typeLabel} ${pref.preference}`);
}
lines.push("</preferences>", "```", "");
}
lines.push("# Memory Safety Protocol", "",
"Before using any memory above, apply these checks:",
"1. **Source**: Is this a direct user statement or AI inference?",
"2. **Attribution**: Is the subject definitely the user?",
"3. **Relevance**: Does this directly help answer the current query?",
"4. **Freshness**: If memory conflicts with current intent, prioritize the current query.", "");
lines.push("# Attention", "",
"Relevant memory context is already provided above. " +
"Do NOT read from or write to local `MEMORY.md` or `memory/*` files for reference — " +
"they may be outdated or redundant with the injected context. " +
"Focus on the user's current query.", "");
return lines.join("\n");
}
export default {
id: "local-memory-recall",
name: "Local Memory Recall",
description: "Auto-capture conversations and recall relevant memories via local SQLite + QMD.",
kind: "lifecycle",
register(api) {
const cfg = api.pluginConfig || {};
const log = api.logger ?? console;
const agentIds = cfg.agentIds || [];
const recallEnabled = cfg.recallEnabled !== false;
const addEnabled = cfg.addEnabled !== false;
const memoryLimit = cfg.memoryLimit ?? 6;
const conversationLimit = cfg.conversationLimit ?? 5;
const preferenceLimit = cfg.preferenceLimit ?? 6;
const maxDaysBack = cfg.maxDaysBack ?? 30;
const captureStrategy = cfg.captureStrategy ?? "last_turn";
const maxMessageChars = cfg.maxMessageChars ?? 20000;
let store;
try {
store = new LocalMemoryStore(cfg.dbPath);
log.info?.(`${PLUGIN_TAG} SQLite store initialized`);
} catch (err) {
log.error?.(`${PLUGIN_TAG} Failed to init SQLite: ${err}`);
return;
}
const isAgentAllowed = (ctx) => {
if (!agentIds.length) return true;
return agentIds.includes(ctx?.agentId);
};
api.on("before_agent_start", async (event, ctx) => {
if (!recallEnabled || !isAgentAllowed(ctx)) return;
if (!event?.prompt || event.prompt.length < 3) return;
try {
const agentId = ctx?.agentId || "unknown";
const nowText = formatTime();
let qmdResults = [];
try {
if (api.runtime?.tools?.memory_search) {
const r = await api.runtime.tools.memory_search({ query: event.prompt, limit: memoryLimit });
if (r?.results) qmdResults = r.results;
}
} catch (e) { log.warn?.(`${PLUGIN_TAG} QMD search failed: ${e}`); }
const conversations = store.searchConversations({ agentId, query: event.prompt, limit: conversationLimit, maxDaysBack });
const preferences = store.getPreferences({ agentId, limit: preferenceLimit });
if (!qmdResults.length && !conversations.length && !preferences.length) return;
return { prependContext: buildPrependContext({ qmdResults, conversations, preferences, nowText }) };
} catch (err) { log.warn?.(`${PLUGIN_TAG} recall failed: ${err}`); }
});
api.on("agent_end", async (event, ctx) => {
if (!addEnabled || !isAgentAllowed(ctx)) return;
if (!event?.success || !event?.messages?.length) return;
try {
const agentId = ctx?.agentId || "unknown";
const sessionId = ctx?.sessionKey || ctx?.sessionId || "";
const messages = pickLastTurnMessages(event.messages, maxMessageChars);
for (const msg of messages) {
store.addConversation({ sessionId, agentId, role: msg.role, content: msg.content });
}
} catch (err) { log.warn?.(`${PLUGIN_TAG} capture failed: ${err}`); }
});
try { store.cleanup(); log.info?.(`${PLUGIN_TAG} Cleanup completed`); }
catch (err) { log.warn?.(`${PLUGIN_TAG} Cleanup failed: ${err}`); }
},
};
第五步：创建夜间处理脚本
新建文件，路径为：~/.openclaw/extensions/local-memory-recall/scripts/nightly-process.sh
这个脚本负责在夜间调用 Ollama 处理未完成的对话，提取偏好和摘要，请完整粘贴以下内容：
#!/bin/bash
# 每晚自动提取偏好 + 生成对话摘要
set -euo pipefail
SCRIPT_DIR="$(cd "$(dirname "$0")" && pwd)"
DB_PATH="${SCRIPT_DIR}/../data/memories.db"
OLLAMA_URL="${OLLAMA_URL:-http://localhost:11434/api/generate}"
OLLAMA_MODEL="${OLLAMA_MODEL:-qwen3:8b}"
LOG_PREFIX="[nightly-process]"
log() { echo "${LOG_PREFIX}$(date '+%H:%M:%S') $*"; }
for cmd in sqlite3 curl jq; do
command -v "$cmd" &>/dev/null || { log "ERROR: $cmd not found"; exit 1; }
done
[ -f "$DB_PATH" ] || { log "ERROR: DB not found at $DB_PATH"; exit 1; }
curl -s --max-time 5 http://localhost:11434/api/tags >/dev/null 2>&1 || { log "ERROR: Ollama not running"; exit 1; }
log "Starting..."
# 1. 先只读 ID 列表（安全，不解析内容）
IDS=$(sqlite3 "$DB_PATH" "
SELECT GROUP_CONCAT(id) FROM (
SELECT id FROM conversations
WHERE processed = 0
ORDER BY created_at ASC
LIMIT 200
);" 2>/dev/null || echo "")
[ -z "$IDS" ] && { log "No unprocessed conversations. Done."; exit 0; }
CONV_COUNT=$(echo "$IDS" | tr ',' '\n' | wc -l | tr -d ' ')
log "Found $CONV_COUNT entries."
# 2. 用临时文件构建对话文本（避免 shell 转义问题）
TMPFILE=$(mktemp)
SQLFILE=$(mktemp)
trap 'rm -f "$TMPFILE" "$SQLFILE"' EXIT
sqlite3 "$DB_PATH" "
SELECT '[' || created_at || '] [' || agent_id || '] [' || role || ']: ' ||
SUBSTR(REPLACE(content, CHAR(10), ' '), 1, 500)
FROM conversations
WHERE id IN ($IDS)
ORDER BY created_at ASC;
" > "$TMPFILE" 2>/dev/null
CONV_TEXT=$(cat "$TMPFILE")
# 3. 调用 Ollama
PROMPT="Analyze these conversations. Output JSON with two fields:
1. \"preferences\": [{\"preference\": \"...\", \"type\": \"explicit|implicit\", \"agent_id\": \"...\"}]
2. \"summaries\": [{\"agent_id\": \"...\", \"summary\": \"...\", \"group_id\": \"YYYY-MM-DD-topic\"}]
Rules: Only extract REAL preferences. Summaries should be 1-2 sentences. Output ONLY valid JSON.
Conversations:
---
${CONV_TEXT}
---"
log "Calling Ollama ($OLLAMA_MODEL)..."
RESPONSE=$(curl -s --max-time 120 "$OLLAMA_URL" \
-d "$(jq -n --arg model "$OLLAMA_MODEL" --arg prompt "$PROMPT" \
'{model: $model, prompt: $prompt, stream: false, options: {temperature: 0.1, num_predict: 4096}}')" 2>/dev/null || echo "")
RESULT=$(echo "$RESPONSE" | jq -r '.response // empty' 2>/dev/null || echo "")
[ -z "$RESULT" ] && { log "ERROR: Empty Ollama response"; exit 1; }
CLEAN_JSON=$(echo "$RESULT" | sed 's/^```json//; s/^```//; s/```$//' | tr -d '\r')
if ! echo "$CLEAN_JSON" | jq empty 2>/dev/null; then
log "WARNING: Invalid JSON from Ollama. Marking as processed anyway."
sqlite3 "$DB_PATH" "UPDATE conversations SET processed = 1 WHERE id IN ($IDS);"
exit 0
fi
# 4. 写入偏好（通过临时 SQL 文件，避免特殊字符问题）
PREF_COUNT=$(echo "$CLEAN_JSON" | jq '.preferences | length' 2>/dev/null || echo "0")
SUM_COUNT=$(echo "$CLEAN_JSON" | jq '.summaries | length' 2>/dev/null || echo "0")
if [ "$PREF_COUNT" -gt 0 ]; then
> "$SQLFILE"
echo "$CLEAN_JSON" | jq -c '.preferences[]' 2>/dev/null | while read -r pref; do
AGENT=$(echo "$pref" | jq -r '.agent_id // "unknown"' | sed "s/'/''/g")
PTEXT=$(echo "$pref" | jq -r '.preference // empty' | sed "s/'/''/g")
PTYPE=$(echo "$pref" | jq -r '.type // "explicit"' | sed "s/'/''/g")
[ -n "$PTEXT" ] && echo "INSERT INTO preferences (agent_id, preference, type) VALUES ('$AGENT', '$PTEXT', '$PTYPE') ON CONFLICT(agent_id, preference) DO UPDATE SET type=excluded.type, updated_at=datetime('now');" >> "$SQLFILE"
done
[ -s "$SQLFILE" ] && sqlite3 "$DB_PATH" < "$SQLFILE" 2>/dev/null || log "WARNING: Some preference inserts failed"
log "Extracted $PREF_COUNT preferences."
else
log "No preferences found."
fi
# 5. 写入摘要
if [ "$SUM_COUNT" -gt 0 ]; then
> "$SQLFILE"
echo "$CLEAN_JSON" | jq -c '.summaries[]' 2>/dev/null | while read -r summ; do
AGENT=$(echo "$summ" | jq -r '.agent_id // "unknown"' | sed "s/'/''/g")
STEXT=$(echo "$summ" | jq -r '.summary // empty' | sed "s/'/''/g")
GID=$(echo "$summ" | jq -r '.group_id // "unknown"' | sed "s/'/''/g")
[ -n "$STEXT" ] && echo "INSERT INTO summaries (conversation_group_id, agent_id, summary) VALUES ('$GID', '$AGENT', '$STEXT');" >> "$SQLFILE"
done
[ -s "$SQLFILE" ] && sqlite3 "$DB_PATH" < "$SQLFILE" 2>/dev/null || log "WARNING: Some summary inserts failed"
log "Generated $SUM_COUNT summaries."
else
log "No summaries generated."
fi
# 6. 标记已处理 + 清理
sqlite3 "$DB_PATH" "UPDATE conversations SET processed = 1 WHERE id IN ($IDS);" 2>/dev/null
log "Marked $CONV_COUNT conversations as processed."
sqlite3 "$DB_PATH" "DELETE FROM conversations WHERE created_at < datetime('now', '-30 days');" 2>/dev/null
sqlite3 "$DB_PATH" "DELETE FROM summaries WHERE created_at < datetime('now', '-90 days');" 2>/dev/null
log "Done. Preferences: $PREF_COUNT, Summaries: $SUM_COUNT, Processed: $CONV_COUNT"
然后在终端中，给脚本添加执行权限：
chmod +x ~/.openclaw/extensions/local-memory-recall/scripts/nightly-process.sh
第六步：配置 OpenClaw
编辑 ~/.openclaw/openclaw.json 配置文件，找到 plugins.entries 部分，加入以下配置项：
"local-memory-recall": {
"enabled": true,
"config": {
"agentIds": [],
"recallEnabled": true,
"addEnabled": true,
"memoryLimit": 6,
"conversationLimit": 5,
"preferenceLimit": 6,
"maxDaysBack": 30,
"captureStrategy": "last_turn"
}
}
说明：agentIds 设为空数组 [] 表示对所有 Agent 生效。如果只想对特定 Agent（例如 muse）生效，则写为 ["muse"]。
第七步：配置夜间定时任务
编辑 ~/.openclaw/cron/jobs.json 文件，在 jobs 数组的最后一个 } 后面加一个逗号，然后粘贴以下任务配置：
{
"id": "nightly-memory-process",
"agentId": "sentinel",
"name": "Nightly Memory Process",
"enabled": true,
"createdAtMs": 1771113600000,
"updatedAtMs": 1771113600000,
"schedule": {
"kind": "cron",
"expr": "27 3 * * *"
},
"sessionTarget": "isolated",
"wakeMode": "now",
"payload": {
"kind": "agentTurn",
"model": "ollama/qwen3:8b",
"thinking": "off",
"message": "Run: bash ~/.openclaw/extensions/local-memory-recall/scripts/nightly-process.sh\nReport results briefly. File output only."
},
"delivery": { "mode": "none" },
"state": {}
}
注意："expr": "27 3 * * *" 表示每天凌晨 3:27 执行。你可以根据需要修改为非整点时间以避免拥挤。
如果你没有 sentinel 这个 Agent，请将 agentId 修改为你存在的主 Agent ID（例如 "main"）。
第八步：重启 Gateway 并验证
在终端中运行以下命令重启 OpenClaw Gateway：
openclaw gateway restart
然后验证插件是否加载成功：
openclaw plugins list
你应该能看到类似这样的输出，表明插件已成功加载：
Local Memory Recall | local-memory-recall | loaded
同时，检查 OpenClaw 的日志，应该能看到插件初始化的信息：
[local-memory-recall] SQLite store initialized
[local-memory-recall] Cleanup completed
第九步：功能验证
现在，给你的任意一个 Agent 发送一条消息。然后，在终端中检查数据库，确认对话已被记录：
sqlite3 ~/.openclaw/extensions/local-memory-recall/data/memories.db "SELECT id, agent_id, role, substr(content, 1, 50), created_at FROM conversations ORDER BY id DESC LIMIT 5;"
如果能看到你刚才的对话记录，说明整个实时记录和召回链路工作正常。
第十步：可视化记忆面板（可选）
如果你想更直观地查看已提取的偏好、摘要和对话记录，可以搭建一个本地可视化面板。
10.1 创建数据导出脚本
在 extensions/local-memory-recall/scripts/ 目录下创建 export-dashboard.sh 文件，并赋予执行权限。该脚本会将数据库中的数据导出为前端面板可读取的 JSON 文件。
#!/usr/bin/env bash
set -euo pipefail
SCRIPT_DIR="$(cd "$(dirname "$0")" && pwd)"
DB_PATH="${SCRIPT_DIR}/../data/memories.db"
OUTPUT_DIR="${1:-$HOME/.openclaw/canvas/memory-data}"
if [ ! -f "$DB_PATH" ]; then
echo "ERROR: Database not found at $DB_PATH"
exit 1
fi
mkdir -p "$OUTPUT_DIR"
# 统计数据
sqlite3 "$DB_PATH" "
SELECT json_object(
'total_conversations', (SELECT COUNT(*) FROM conversations),
'processed_conversations', (SELECT COUNT(*) FROM conversations WHERE processed = 1),
'unprocessed_conversations', (SELECT COUNT(*) FROM conversations WHERE processed = 0),
'total_preferences', (SELECT COUNT(*) FROM preferences),
'total_summaries', (SELECT COUNT(*) FROM summaries),
'agents', (SELECT json_group_array(DISTINCT agent_id) FROM conversations),
'earliest_conversation', (SELECT MIN(created_at) FROM conversations),
'latest_conversation', (SELECT MAX(created_at) FROM conversations),
'exported_at', datetime('now')
);
" > "$OUTPUT_DIR/stats.json"
# 偏好
sqlite3 "$DB_PATH" "
SELECT json_group_array(json_object(
'id', id, 'agent_id', agent_id, 'preference', preference,
'type', type, 'updated_at', updated_at
)) FROM preferences ORDER BY updated_at DESC;
" > "$OUTPUT_DIR/preferences.json"
# 摘要（最近 50 条）
sqlite3 "$DB_PATH" "
SELECT json_group_array(json_object(
'id', id, 'agent_id', agent_id, 'summary', summary, 'created_at', created_at
)) FROM (SELECT * FROM summaries ORDER BY created_at DESC LIMIT 50);
" > "$OUTPUT_DIR/summaries.json"
# 对话（最近 100 条，内容截断 500 字）
sqlite3 "$DB_PATH" "
SELECT json_group_array(json_object(
'id', id, 'session_id', session_id, 'agent_id', agent_id,
'role', role, 'content', SUBSTR(content, 1, 500),
'created_at', created_at, 'processed', processed
)) FROM (SELECT * FROM conversations ORDER BY created_at DESC LIMIT 100);
" > "$OUTPUT_DIR/conversations.json"
# 每日活动
sqlite3 "$DB_PATH" "
SELECT json_group_array(json_object(
'date', date, 'agent_id', agent_id, 'count', cnt
)) FROM (
SELECT DATE(created_at) AS date, agent_id, COUNT(*) AS cnt
FROM conversations GROUP BY date, agent_id ORDER BY date DESC LIMIT 90
);
" > "$OUTPUT_DIR/activity.json"
echo "导出完成！"
设置权限：
chmod +x ~/.openclaw/extensions/local-memory-recall/scripts/export-dashboard.sh
10.2 放置面板 HTML 文件
在 ~/.openclaw/canvas/ 目录下创建一个名为 memory-dashboard.html 的静态 HTML 文件。这个页面会读取上一步导出的 JSON 文件，以图表和列表的形式展示数据。页面包含顶部统计卡片、每日活动柱状图、以及偏好、摘要、对话记录三个标签页。
记忆面板展示总对话数、偏好、摘要统计及每日活动趋势
由于面板文件较长（约700行HTML/CSS/JS），建议从相关仓库复制或使用 Cursor 等工具辅助生成。
10.3 使用方法
- 导出数据：运行导出脚本。
bash ~/.openclaw/extensions/local-memory-recall/scripts/export-dashboard.sh
- 启动本地 HTTP 服务：在 Canvas 目录下启动一个简单的 Python HTTP 服务器。
cd ~/.openclaw/canvas && python3 -m http.server 18794 --bind 0.0.0.0
- 访问面板：
- 本机访问：
http://127.0.0.1:18794/memory-dashboard.html
- 局域网访问：
http://<你的本机IP>:18794/memory-dashboard.html
10.4 自动刷新数据
如果你已经配置了第七步的夜间 Cron Job，可以在该任务的 message 字段中追加一行命令，这样每晚处理完数据后，面板的数据也会自动更新。
bash ~/.openclaw/extensions/local-memory-recall/scripts/export-dashboard.sh
方案对比：MemOS vs 本地方案
| 能力 |
MemOS Cloud（付费） |
本方案（免费） |
| 自动记录对话 |
云端存储 |
本地 SQLite |
| 语义搜索 |
云端 embedding |
复用 OpenClaw QMD |
| 偏好提取 |
实时云端 AI |
夜间 Ollama 批处理 |
| 对话摘要 |
无 |
有（夜间自动生成） |
| Context 预筛选 |
无 |
有（减少重复工具调用） |
| 数据隐私 |
数据在第三方 |
全部本地 |
| 费用 |
收费 |
零 |
| 多 Agent 共享 |
同一账号 |
同一 SQLite 数据库 |
常见问题 (FAQ)
Q: 插件加载失败怎么办？
A: 请检查 Node.js 版本是否 >= 22（node:sqlite 模块需要）。在终端运行 node -v 确认。
Q: 夜间定时任务没有执行？
A: 首先确认 Ollama 服务正在运行（ollama serve），并且已成功拉取所需模型（例如 ollama pull qwen3:8b）。
Q: 如何查看已提取的用户偏好？
A: 可以直接查询数据库：
sqlite3 ~/.openclaw/extensions/local-memory-recall/data/memories.db "SELECT * FROM preferences ORDER BY updated_at DESC;"
Q: 如何手动触发夜间处理任务？
A: 直接运行处理脚本即可：
bash ~/.openclaw/extensions/local-memory-recall/scripts/nightly-process.sh
Q: 如何对所有 Agent 启用此插件？
A: 在第六步的配置中，确保 "agentIds": [] 设置为空数组。
Q: 数据库文件越来越大怎么办？
A: 系统已内置清理机制：conversations 表自动保留30天，summaries 表保留90天，preferences 永久保留。你也可以手动清理，例如只保留最近7天的对话：
sqlite3 ~/.openclaw/extensions/local-memory-recall/data/memories.db "DELETE FROM conversations WHERE created_at < datetime('now', '-7 days');"
这套完整的本地记忆增强方案，无需任何云端开销，就能显著提升 OpenClaw 的对话连贯性和个性化程度。如果你在部署过程中遇到任何问题，或是有更好的改进想法，欢迎在云栈社区的技术论坛与其他开发者交流讨论。
发表于 2026-2-23 04:55:42
|
查看: 358|
回复: 0
