MemTree - Memory for long-running agents

How It Works

We compress older messages with our state-of-the-art MemTree algorithm to keep AI fast and intelligent

📖 Technical Details Deep dive into the architecture Video See our video explainer

Quickstart

Claude Code Infinite

Install:

npm install -g claude-code-infinite

Run:

ccc

Recommended: Read the full docs here

In Settings > Providers

Set Base URL to
https://polychat.co/api
Paste in your PolyChat API key
Choose a memtree- model

Optional Settings

In Settings > Providers

Set the Context Window Size to your model's limit, i.e. 200000 for memtree-sonnet-4-thinking-low

In Settings > Context

At the bottom, uncheck "Automatically trigger intelligent context condensing" - as we do this automatically

Get your PolyChat API key

import openai

client = openai.OpenAI(
    base_url="https://polychat.co/api",
    api_key="YOUR_POLYCHAT_API_KEY"
)

response = client.chat.completions.create(
    model="memtree-sonnet-4",
    messages=[{"role": "user", "content": "Hello!"}],
)

print(response.choices[0].message.content)

📖 Full Documentation

⚡ BREAKTHROUGH PERFORMANCE

Long-running task performance

Context Memory changes everything for AI reasoning tasks

Accuracy advantage of Context Memory

Accuracy

1.0 0.8 0.6 0.4 0.2 0.0

0.994

With memory

0.690

W/out memory

44% improvement

📊 memtree.dev

🏆

Industry First Achievement

We used Context Memory to crack one of @METR_Evals hardest evals: an ML task which no other model has previously solved