據 AlphaMoat 統計,截至 2026-09,Vanar Neutron Memory 累計下載 3,311 次、獲 1 個星標,在全部 5,078 個 AI 技能中排名第 3,415 位,在「AI 智慧體」分類(969 個)中排名第 701 位。
| 平台 | 累計安裝 | 累計下載 | 星標 | 評論 |
|---|---|---|---|---|
| ClawHub ↗ | 101 | 3158 | 1 | - |
| SkillHub ↗ | 297 | 3311 | 1 | — |
各平台的計數口徑不同,統計時間不一致。
它把對話、偏好與決策跨會話持久化,支援手動 save/search 命令,也可選擇開啟 Auto-Capture 與 Auto-Recall 鉤子,在每輪 AI 前注入相關記憶。資料透過 HTTPS 發往 Neutron API,文件說明預設不傳送任何資料,僅在執行命令或顯式開啟鉤子時才上傳明確儲存的內容。
資料釋出於 2026-10-27
以下是根據技能文件生成的示例,不代表實際執行結果。
2026-09 起統計方法調整:同名技能按作者分別統計,並覆蓋全部技能。圖中虛線為調整前的月份,與之後的資料不直接比較,也不計算跨越調整的增長。
| 月份 | 累計下載 | 環比 | 星標 |
|---|---|---|---|
| 2026-09 | 3311 | — | 1 |
| 2026-08 | 3049 | +4.3% | 1 |
| 2026-07 | 2922 | +8.9% | 1 |
| 2026-06 | 2683 | +5.3% | 1 |
| 2026-05 | 2549 | — | 1 |
| 2026-03 | 2277 | — | 1 |
以下是作者釋出的原始檔案,按純文字顯示,本站未做修改或稽核;其中的連結不可點選,命令請核實後再執行。抓取於 2026-10-23。
---
name: vanar-neutron-memory
description: Save and recall agent memory with semantic search. Context that persists across every session.
user-invocable: true
metadata: {"openclaw": {"emoji": "🧠", "requires": {"env": ["API_KEY"]}, "primaryEnv": "API_KEY"}}
---
# Neutron Memory
Every conversation, preference, and decision your agent makes can persist across sessions. Save what matters, and when you need it, semantic search finds the right context by meaning — not keywords. Every session builds on the last.
## How It Works
**Manual** — save and search with simple commands:
1. `./scripts/neutron-memory.sh save "user prefers dark mode" "Preferences"` — save context
2. `./scripts/neutron-memory.sh search "what theme does the user like"` — find it by meaning
**Automatic** (opt-in) — enable hooks to capture and recall automatically:
1. **Auto-Capture** saves conversations after each AI turn
2. **Auto-Recall** finds relevant memories before each AI turn and injects them as context
## Quick Start
See **[SETUP.md](SETUP.md)** for the full setup guide. TL;DR:
1. Get a free API key at **https://openclaw.vanarchain.com/** ($20 free credits, no credit card)
2. `export API_KEY=nk_your_key`
3. `./scripts/neutron-memory.sh test`
## Commands
### Save
```bash
./scripts/neutron-memory.sh save "Content to remember" "Title"
```
### Search
```bash
./scripts/neutron-memory.sh search "what do I know about X" 10 0.5
```
Arguments: `QUERY` `LIMIT` `THRESHOLD(0-1)`
### Diagnose
```bash
./scripts/neutron-memory.sh diagnose
```
Checks all prerequisites: curl, jq, API key, connectivity, and authentication.
## Hooks (Auto-Capture & Auto-Recall)
- `hooks/pre-tool-use.sh` — **Auto-Recall**: Queries memories before AI turn, injects relevant context
- `hooks/post-tool-use.sh` — **Auto-Capture**: Saves conversation after AI turn
Both are **disabled by default** (opt-in only). To enable:
```bash
export VANAR_AUTO_RECALL=true
export VANAR_AUTO_CAPTURE=true
```
## API Endpoints
- `POST /memory/save` — Save text (multipart/form-data)
- `POST /memory/search` — Semantic search (JSON body)
**Auth:** `Authorization: Bearer $API_KEY` — that's it. No other credentials needed.
## Security & Privacy
**No data is sent unless you run a command or explicitly enable auto-capture/auto-recall.** Both hooks are disabled by default.
This skill only sends data you explicitly save (or opt-in auto-captured conversations) to the Neutron API. Here's exactly what happens:
| Action | What's sent | Where |
|--------|------------|-------|
| **Save** | The text you pass to `save` or auto-captured conversation turns | `POST /memory/save` over HTTPS |
| **Search** | Your search query text | `POST /memory/search` over HTTPS |
| **Auto-Recall** | The user's latest message (used as search query) | `POST /memory/search` over HTTPS |
| **Auto-Capture** | `User: {message}\nAssistant: {response}` | `POST /memory/save` over HTTPS |
**What is NOT sent:**
- No local files are read or uploaded
- No environment variables (other than the API key for auth)
- No system information, file paths, or directory contents
- No data from other tools or skills
All communication is over HTTPS to `api-neutron.vanarchain.com`. The source code is fully readable in the `scripts/` and `hooks/` directories — three short bash scripts, no compiled binaries.
將開發過程中的經驗、錯誤與修正記錄為 Markdown 檔案,供日後重用並提升為工作區記憶。
用自然語言描述需求,在本機與遠端技能市場尋找技能並取得推薦。
給 AI agent 用的瀏覽器自動化命令列技能。
將使用者明確糾正與自我檢視整理成分層記憶,並依使用情況升級或封存。
在安裝第三方技能之前,按清單審查來源、許可權與危險模式的技能。
將代理記憶與跨技能狀態整理成具型別與約束的知識圖譜。
規劃 AI 代理主動跟進、工作記憶與回顧流程的工作區設定指南。
指導如何建立與更新有效 Claude 技能的寫作指南。
資料月份:2026-09 · 下載/星標/安裝數聚合自公開技能註冊庫(ClawHub、SkillHub),按月更新。口徑說明見資料方法。