Memok AI — OpenClaw plugin

English | 简体中文 · Website: memok-ai.com · Plugin mirror: Gitee

This repository is the OpenClaw gateway extension for Memok memory. The npm package name for this repo is memok-ai-openclaw. After openclaw plugins install, OpenClaw puts the extension under ~/.openclaw/extensions/<plugin id>/, where <plugin id> comes from openclaw.plugin.json → id (currently memok-ai), not from package.json name — so the live folder is ~/.openclaw/extensions/memok-ai/. The full memory engine is the separate npm package memok-ai; details below after install.

Requirements

• Node.js ≥20 (LTS recommended), npm
• OpenClaw gateway ≥2026.3.24 (see openclaw.compat in package.json)

Install (OpenClaw plugin)

One-line install: paste into a terminal. curl -fsSL downloads the script from the raw URL; bash <(...) runs it in one step (process substitution—no separate “save file then chmod” step). You need git, node, npm, and openclaw already on your PATH.

Linux / macOS — default (npm registry for memok-ai)

scripts/install-linux-macos.sh — installs core from npm memok-ai ("memok-ai": "^0.2.2" in package.json; imports use memok-ai/bridge).

bash <(curl -fsSL https://raw.githubusercontent.com/galaxy8691/memok-ai-openclaw/main/scripts/install-linux-macos.sh)

Linux / macOS — mainland China (Gitee clone + npmmirror)

scripts/install-cn-linux-macos.sh — clones the plugin from Gitee when possible, then npm install --registry (default npmmirror) so memok-ai is fetched from the mirror. Optional: set MEMOK_CORE_GIT_URL to force memok-ai from Git instead of npm.

bash <(curl -fsSL https://gitee.com/wik20/memok-ai-openclaw/raw/main/scripts/install-cn-linux-macos.sh)

If registry.npmjs.org is slow or blocked, use the China line (npmmirror) or set MEMOK_NPM_REGISTRY before install-linux-macos.sh. If the memok-ai tarball is unavailable on your registry, set MEMOK_CORE_GIT_URL (optional MEMOK_CORE_GIT_REF) to install the core from Git.

If you already cloned this repo:

bash scripts/install-linux-macos.sh          # from repo root — same as curl one-liner
# or
bash scripts/install-cn-linux-macos.sh

Windows

scripts/install-windows.ps1 — irm downloads the script; | iex runs it.

Default (GitHub raw; core from npm):

irm https://raw.githubusercontent.com/galaxy8691/memok-ai-openclaw/main/scripts/install-windows.ps1 | iex

Clone plugin from Gitee (installer uses npmmirror for npm install when the clone URL is Gitee):

$env:MEMOK_REPO_URL = "https://gitee.com/wik20/memok-ai-openclaw.git"
irm https://gitee.com/wik20/memok-ai-openclaw/raw/main/scripts/install-windows.ps1 | iex

What the installers do: npm install → npm run build → openclaw plugins install → openclaw memok setup → optional gateway restart → remove ~/.openclaw/extensions/memok-ai-openclaw-src unless MEMOK_KEEP_SOURCE=1.

Useful env vars: MEMOK_PLUGINS_INSTALL_TIMEOUT_SECONDS, MEMOK_PLUGINS_INSTALL_NO_PTY=1, MEMOK_SKIP_GATEWAY_RESTART=1, MEMOK_GATEWAY_RESTART_TIMEOUT_SECONDS, MEMOK_KEEP_SOURCE=1, MEMOK_REPO_URL / MEMOK_REPO_URL_CN / MEMOK_REPO_URL_FALLBACK, MEMOK_CORE_GIT_URL, MEMOK_CORE_GIT_REF, MEMOK_NPM_REGISTRY (cn script defaults to npmmirror) — see script headers and README.zh-CN.md for Chinese details.

If plugins.allow blocks memok, add "memok" under plugins.allow in ~/.openclaw/openclaw.json, then run openclaw memok setup again.

Manual install (no curl script)

From GitHub:

git clone https://github.com/galaxy8691/memok-ai-openclaw.git
cd memok-ai-openclaw
npm install && npm run build
openclaw plugins install .
openclaw memok setup

From Gitee (use a China-friendly registry so memok-ai resolves quickly):

git clone https://gitee.com/wik20/memok-ai-openclaw.git
cd memok-ai-openclaw
npm install --registry https://registry.npmmirror.com && npm run build
openclaw plugins install .
openclaw memok setup

If you cannot use npm for the core package, before npm install:
npm pkg set dependencies.memok-ai=git+https://gitee.com/wik20/memok-ai.git#v0.1.0

Air-gapped: build core locally, then set "memok-ai": "file:/path/to/memok-ai" in package.json and npm install.

Core vs plugin (repos)

	Repository	Role
Core	galaxy8691/memok-ai · Gitee mirror	Pipelines, CLI, tests; npm memok-ai, import memok-ai/bridge.
Plugin (this repo)	galaxy8691/memok-ai-openclaw · Gitee mirror	src/plugin.ts, openclaw.plugin.json, skills/ only.

Single package.json: dependency memok-ai at ^0.2.2 (memok-ai on npm; plugin code imports memok-ai/bridge). First npm install runs that package’s prepare → npm run build (includes native better-sqlite3, often minutes on a cold cache).

Memok pipeline config: after openclaw memok setup, the plugin writes ~/.openclaw/extensions/memok-ai/config.toml (a MemokPipelineConfig). At gateway startup the plugin loads only that file for articleWordPipeline / dreaming / recall; if it is missing, the plugin logs an error and skips registration until you run setup again. This is separate from openclaw.json (which still holds cron, recall UI flags, etc.).

China: use install-cn-linux-macos.sh (npmmirror) or MEMOK_NPM_REGISTRY. Git fallback for core: set MEMOK_CORE_GIT_URL (and optional MEMOK_CORE_GIT_REF) before npm install.

Development (clone this repo)

git clone https://github.com/galaxy8691/memok-ai-openclaw.git
cd memok-ai-openclaw
npm install   # installs memok-ai from npm
npm run build
npm run ci    # lint + build + tests

If you cannot pull memok-ai from npm, before npm install:

npm pkg set dependencies.memok-ai=git+https://gitee.com/wik20/memok-ai.git#v0.1.0
npm install

What the plugin does

• Persists transcripts, per-turn recall (before_prompt_build), tools memok_recall_candidate_memories / memok_report_used_memory_ids, optional dreaming cron
• openclaw memok setup wizard and plugin config

Evaluation (tested): with recall + report flow, effective use of candidate memories exceeded 95% in our runs; your mileage varies by model and settings.

Compared to embedding-only stacks

	Memok	Typical hosted vector DB
Deployment	Local SQLite	Cloud API + billing
Recall	Word graph, weights, sampling	Embedding similarity
Explainability	Inspectable rows	Mostly scores

Notes from real use

Users report coherent multi-session follow-up and stable feedback when citing memories; DBs with ~1k sentences / 100k+ link rows are not uncommon. Timing is workload-dependent (~10² ms class for persist in informal local tests), not a SLA.

CLI / pipelines / dreaming one-shots

Use the core repo: README · README.zh-CN.

Dreaming (plugin cron)

When dreaming cron runs, each run is stored in SQLite dream_logs: dream_date, ts, status (ok / error), log_json.

Config priority (OPENAI_*, MEMOK_LLM_MODEL)

1. Existing environment variables win
2. Plugin config fills gaps only
3. For this plugin, LLM + SQLite pipeline settings come from ~/.openclaw/extensions/memok-ai/config.toml (see openclaw memok setup). Core CLI may still document its own dev env elsewhere.

Plugin users normally rely on openclaw memok setup only.

Contributing

See CONTRIBUTING.md.

License

MIT License.