ChronoLLM

Chat with temporally consistent language models. Each model only knows events up to its vintage year — no future knowledge leaking into your analysis.

Install

pip install chronollm

Setup

# Save your HuggingFace token (provided by the ChronoLLM team)
chronollm login <your_token>

Usage

# Chat with a vintage year
chronollm 2020

# Chat with a specific model
chronollm owner/repo

Example

> What caused the 2020 pandemic?
  The 2020 pandemic was caused by the novel coronavirus SARS-CoV-2,
  which emerged in late 2019 and spread globally.

> How did governments respond?
  Governments responded with lockdowns, travel restrictions, economic
  stimulus packages, and accelerated vaccine development programs.

> /clear
  Context cleared.

> Compare solar and wind energy
  Solar energy is more predictable and scalable for rooftop installations,
  while wind energy is more efficient at large scale but depends on location
  and weather patterns.

Requirements

• Python 3.10+
• Works on CUDA (GPU), Apple Silicon (MPS), and CPU
• ~4GB disk space per model, ~8GB RAM

Tips

• Be specific: "What was the state of AI in 2018?" works better than "Tell me about AI"
• Include the year: "Summarize the 2015 housing market" gives better results
• Use tasks: summarize, compare, explain, list, evaluate
• Paste text directly after your instruction, don't wrap in quotes
• Standalone questions give more accurate answers than follow-ups

Limitations

• 2B parameter model: understands topics well but may invent specific names, dates, or numbers
• Context window is ~2048 tokens (~5-10 turns of conversation)
• Always verify facts for critical use cases

Commands

Development

git clone git@github.com:chronollm/cli.git
cd cli
uv sync
uv run chronollm 2020