360 Secure SkillHub
360 安全技能中心

Memkeeper for OpenClaw

面向 OpenClaw 用户的 Memkeeper 本地记忆配置 Skill:下载并初始化本地嵌入/重排模型与 SQLite 存储,通过 OpenClaw MCP 命令注册受限的 12 工具配置,运行 doctor 探针验证连接,并可按要求卸载;默认保持记忆文本本地处理,仅在选择远程嵌入或重排服务商时才会外发文本。

teflon07
teflon07
热度 7.64
最近更新 2026-07-10 21:31:53
分类:知识依赖:代码执行 / 读写文件 / MCP 服务应用场景:笔记整理 / 工作流

安装技能

复制下面的话术发给你的 Agent,它会通过 secure-skills 检索并安装该技能。

执行命令安装此 skill: npx secure-skills-cli add memkeeper-mcp-setup
Metadata
name
memkeeper-openclaw
description
Install, configure, verify, or remove local-first Memkeeper memory for OpenClaw. Use when an OpenClaw user needs Memkeeper's local embedding and reranking models, a safe MCP tool profile, or diagnosis of a failed Memkeeper MCP connection.

Memkeeper for OpenClaw

Configure Memkeeper as a local stdio MCP server. Keep embedding and reranking local by default. Do not select an OpenClaw reasoning model, configure a model provider, install a Gateway service, or enable a messaging channel unless the user explicitly asks.

Set up local semantic memory

Use a user-level store unless the user requests project isolation.

MEMKEEPER_BIN="${MEMKEEPER_BIN:-$HOME/.local/bin/memkeeper}"
MEMKEEPER_STORE="${MEMKEEPER_STORE:-$HOME/.memkeeper/store.sqlite}"

If MEMKEEPER_BIN is absent, download the public installer to a temporary file, inspect it, and run it. Do not paste secrets into a command or config file.

curl -fsSL https://raw.githubusercontent.com/teflon07/memkeeper/main/install.sh \
  -o /tmp/memkeeper-install.sh
bash /tmp/memkeeper-install.sh

Download the local models, initialize the store, and require semantic retrieval:

"$MEMKEEPER_BIN" pull-models
"$MEMKEEPER_BIN" init --store "$MEMKEEPER_STORE" --json
MEMKEEPER_REQUIRE_SEMANTIC=1 \
  "$MEMKEEPER_BIN" doctor --store "$MEMKEEPER_STORE" --json

Require the final doctor output to report semantic models as ready. Treat a missing model as a setup failure, not as permission to use lexical fallback.

Register the MCP server

Register Memkeeper through OpenClaw's MCP command. Expose the tested 12-tool profile first:

openclaw mcp add memkeeper \
  --command "$MEMKEEPER_BIN" \
  --arg mcp \
  --env "MEMKEEPER_STORE=$MEMKEEPER_STORE" \
  --env "MEMKEEPER_REQUIRE_SEMANTIC=1" \
  --include "stats,search,get,memory_list,entity_search,graph_neighbors,graph_context,dream_graph,remember,pack,candidate_submit,candidate_list" \
  --timeout 30 \
  --connect-timeout 30

Keep forget, entity_upsert, relationship_upsert, and verify disabled until the operator explicitly needs deletion or direct graph/metadata mutation.

Verify and report

Prove the live stdio connection after registration:

openclaw mcp doctor memkeeper --probe --json

Report the configured store path, enabled tool count, and probe result. Do not print stored memory contents, API keys, tokens, or other sensitive environment values.

The MCP connection test does not require an OpenClaw reasoning provider, a Gateway daemon, or messaging channels. Those are independent, user-selected choices.

Handle non-local embedding only on request

Do not change the local mxBai embedding and reranking defaults without explicit operator direction. An API embedding or reranking provider receives memory text outside the machine and changes the store's vector backend. Use a separate store or deliberately re-embed before switching.

For removal, unregister the MCP server with openclaw mcp unset memkeeper. Leave the local store and models in place unless the user explicitly requests their removal.