One memory for all your local AI.
RAMen sits between your AI apps and your models and gives all of them memory. Every chat and agent session comes back in seconds, long ones stay fast, and they follow you to your phone.
Free plan forever ✦ 14 days of everything, no card ✦ Mac, Windows and Linux, including ARM
Launching on Product Hunt this Sunday, Nov 8. Get a reminder
Keep your apps and models. RAMen adds the memory.
Apps made for Ollama, OpenAI or Anthropic talk to RAMen without a change. Every model on your computer is one pick away, and each conversation keeps its memory wherever it goes.
Your apps
Open WebUI
Continue
Cline
Aider
OpenCode
Claude Code
Codex
and any app made for Ollama or OpenAI
RAMen
memory for every chat, agent and model
Your models
Ollama
LM Studio
Hugging Face
Your phone
RAMen app
Already use Ollama?ramen start --ollama replace
Claude Coderamen claude
Codexramen connect codex
Long chats stop slowing down.
The longer a conversation gets, the slower a local model writes. RAMen reads only the parts of the conversation that matter at each step, so the speed holds.
Measured on an NVIDIA L4 with Qwen3-4B, November 2026: 42 versus 9.1 tokens per second at 128K tokens, same settings. The video: Qwen3 32B on an RTX A6000, 20.5 versus 8.4 tokens per second at 87K tokens, same answer. How it works and all results.
Three commands. No new tools.
RAMen picks a model that fits your computer, sizes the memory, and speaks the same APIs as Ollama, OpenAI and Anthropic, so your apps work unchanged.
STEP 01Install
One line. Mac (Apple silicon) and Linux, in Terminal:
curl -fsSL https://getramenai.com/install.sh | bashWindows 10 and 11, in PowerShell:
irm https://getramenai.com/install.ps1 | iex- STEP 02
Start
It finds a model that fits, downloads it and starts in the background. Or start one you already downloaded with Ollama or LM Studio, or any GGUF file.
ramen start 
STEP 03Connect your apps
Apps made for Ollama find RAMen at Ollama's usual address (if you use Ollama,
ramen start --ollama replacetakes its place). Cline, Continue, Aider, OpenCode, Open WebUI and Codex in one command; Claude Code withramen claude.ramen connect cline
Memory that keeps up with your agents.
Sessions that survive restarts
Every conversation is saved to disk on its own and comes back in seconds, even after a reboot.
Many agents, one GPU
Switch between agents and chats without re-reading their history. Each one resumes where it stopped.
Fast on long chats
Selective reading switches on by itself where it is faster on your machine, so RAMen is never slower.
Models bigger than your memory
Run a mixture-of-experts model larger than your computer's memory, like gpt-oss 120B on a 16 GB Mac. RAMen reads its experts straight from the file into its own cache. On Pro.
Works with your apps and models
One command connects Cline, Continue, Aider, OpenCode, Open WebUI, Claude Code and Codex, and apps made for Ollama work too. Every model already on your computer shows up in them, and picking one switches RAMen to it.
Stays on your computer
Prompts, files and sessions never leave your machine, and it works offline. Built on llama.cpp, the engine local AI already runs on.
Measured on a Mac mini M2 Pro with 16 GB and gpt-oss 120B (63 GB), November 2026, same settings: a 1,233-token prompt read at 62 versus 14 tokens per second by plain llama.cpp, and the model started in 5 seconds versus 41. Writing runs at 2.8 tokens per second, up from 1.9 with RAMen 1.0. Big models are part of Pro and the trial. How it works.
Start on your computer. Continue on your phone.
Hand a conversation to your phone with its memory already read. The RAMen app answers your next message right away, on the phone and offline. Nothing goes through a cloud.
- STEP 01
Pack it on your computer
Your computer reads the whole conversation with the phone app's model and puts that memory in one file.
ramen handoff - STEP 02
Send it to your phone
AirDrop to an iPhone, Quick Share or a cable to an Android phone. The file goes straight from your computer to your phone.
- STEP 03
Keep talking
Open it with the RAMen app. Your phone loads the memory instead of reading the whole conversation again, so it answers right away.
The phone app runs a small model (Qwen3 0.6B) on the phone itself and saves every chat's memory, so chats come back in about a second even after the phone closes the app. Android 10 or newer on a 64-bit phone: open the download on the phone and allow the install once. The iPhone app is in testing.
Free forever. Faster for $3.99.
Start free. Every plan begins with 14 days of everything, no card needed.
Free
- Remembers your last 3 chats, back in seconds
- One chat at a time
- Works with your apps, one command to start
Pro Popular
- Remembers every chat and agent session
- Many agents at once on one GPU
- Long chats stay fast, plus the draft helper on supported models
- Models bigger than your computer's memory
- Up to 3 of your computers
Team
- Pro for everyone on the team
- One bill, seats you can move
- Priority email support
Server
- One GPU machine serving your whole team
- A key per person, conversations kept apart
- Usage per person and Prometheus metrics
Enterprise
Offline licenses for machines without internet, invoices and purchase orders, a support agreement.
Secure checkout by Polar ✦ Sales tax handled at checkout ✦ Cancel any time in your customer portal
Good to know.
What do I need to run it?
A Mac with Apple silicon, or a Windows or Linux PC, with or without an NVIDIA card; Linux on ARM works too (for example NVIDIA DGX Spark or AWS Graviton). RAMen picks a model that fits, or use any GGUF model, including the ones you already downloaded with Ollama or LM Studio. Selective reading on long chats needs an NVIDIA RTX 20-series card or newer, or an Apple Silicon Mac. Session memory works everywhere.
Can it run a model bigger than my computer's memory?
Yes, on Pro and during the trial, for mixture-of-experts models such as gpt-oss 120B or Qwen3 30B-A3B. RAMen keeps the rest of the model in memory and reads the experts from the SSD as they are needed, into its own cache. It writes more slowly than a model that fits, and the answers are the same. ramen models shows what runs on your computer.
Does it work with Ollama or LM Studio?
Yes. The models you downloaded with Ollama or LM Studio show up in your apps and in ramen models, and start without a new download. Apps made for Ollama can talk to RAMen, and ramen start --ollama replace puts RAMen in Ollama's place, so they use it without a change. ramen stop starts Ollama again.
Does it work on phones?
Yes. The RAMen app for Android (preview) runs a small model on the phone with saved memory for every chat; the iPhone app is in testing. ramen handoff moves a conversation from your computer to your phone with its memory already read, so the phone can continue it right away.
Does it send my data anywhere?
No. Everything runs on your computer. Paid plans check the license key at start-up, and RAMen keeps working offline between checks.
Which apps work with it?
Anything that talks to the OpenAI API, Anthropic's API or Ollama's API. ramen connect sets up Cline, Continue, Aider, OpenCode, Open WebUI and Codex for you, and ramen claude starts Claude Code with RAMen's model.
Can it make Claude or ChatGPT faster?
No. Those models run on Anthropic's and OpenAI's own servers. RAMen speeds up the models on your own computer, and it can run Claude Code and Codex, the coding tools, on those models with their sessions saved.
What happens after the trial?
RAMen keeps working on the free plan: it remembers your last 3 chats, one at a time. Add a Pro key with ramen activate YOUR-KEY to remember everything, run many agents at once and keep long chats fast.
Is there a one-time price?
Yes, at launch: Pro for life for $19.99, for the first 1,000 buyers. After that Pro is $3.99 a month or $29 a year.
Who is the Server plan for?
Teams that share one GPU machine. RAMen serves everyone from it, gives each person a key, keeps each person's conversations apart and counts usage per person.
How do I switch on my plan after buying?
Your license key is in the email from Polar and in your customer portal. Run ramen activate YOUR-KEY and you are done.
How do I cancel or change my plan?
Any time, in your customer portal. You keep your plan until the end of the period you paid for, then RAMen goes back to the free plan. See the refund policy.
Can I move my license to a new computer?
Yes. A Pro key works on three computers at a time, and you can free a slot yourself with ramen deactivate or from the customer portal.