memory-keeper

Stop paying for empty context. Keep 99.9% DeepSeek prompt cache hit rate.

🎯 99.95% cache hit in production

📊 The Problem

❌ Without memory-keeper

14%

cache hit rate
cacheMiss: 101,295

✅ With memory-keeper

99.95%

cache hit rate
cacheMiss: 31

🏗️ How It Works

System (stable) └── Active (15 rounds, full) └── Archive (30 rounds, in-place compression → 100% cache hit) └── Super Archive (batch compressed → 1 round recovery)

A lightweight Node.js proxy between OpenClaw and DeepSeek API. 3-tier message archiving keeps the prefix stable — normal ops never miss cache.

⚡ Key Features

100%

Normal cache hit

Round recovery

5 min

Setup time

Config changes

🛠️ Tech Stack

Node.js · Express · 3-tier archive · BUSL-1.1 (→ Apache 2.0 in 2029)

⭐ GitHub Install on ClawHub