RAMen

RAMen is a memory layer for local language models running on llama.cpp. Key features: • Saves every chat and agent session on your computer • Reopens long sessions in seconds, even after a restart • Reads only the relevant parts of long conversations to keep replies fast • Shares one GPU across several agents without losing their history • One-command install on Mac (Apple silicon), Windows and Linux • Works offline with the apps you already use The longer a conversation grows, the slower a local model typically responds. RAMen addresses this by processing only the parts of the conversation that matter at each step, so speed holds as the chat grows. In the published benchmarks, a 105,000-token session reopened in 1.8 seconds instead of 3.7 minutes, and replies at 128,000 tokens ran about 5.5 times faster. Setup takes a single command. RAMen selects a model that fits your hardware, sizes memory accordingly, and exposes the same API as OpenAI, so existing apps connect without changes. Everything stays on your machine. RAMen suits developers, researchers, and power users who run local models for long writing sessions, document analysis, or multi-agent experiments. The free plan covers session memory permanently.