A Python tool that uses rolling summarization to condense long chat conversations while preserving narrative flow, emotional context, and key information.

• Rolling Summarization: Processes conversations in chunks, with each summary building on the previous one
• Preserves Context: Maintains emotional journey, key decisions, and important relationships
• Configurable: Customizable prompts and chunk sizes
• Debug Mode: Save intermediate summaries to track the summarization process
• Progress Tracking: Visual progress bar for long conversations

• Python 3.11+
• OpenAI API key

1. Clone the repository:

git clone https://github.com/yourusername/llm-chat-summarizer.git
cd llm-chat-summarizer

2. Install dependencies using uv (recommended):

pip install uv
uv sync

Or using pip with pyproject.toml:

pip install -e .

3. Set your OpenAI API key:

export OPENAI_API_KEY="your-api-key-here"

uv run python main.py chat_export.md --seed-prompt seed_prompt.txt

uv run python main.py chat_export.md --seed-prompt seed_prompt.txt --summary-prompt custom_prompt.txt

uv run python main.py chat_export.md --seed-prompt seed_prompt.txt --max-chunks 5 --debug

• input_file: Path to the markdown file containing the chat conversation (required)
• --seed-prompt: Path to the seed prompt file (required)
• --summary-prompt: Path to custom summarization prompt file (optional)
• --output: Output file path (default: final_summary.txt)
• --max-chunks: Process only first N chunks (useful for testing)
• --debug: Save intermediate summaries for each chunk

1. Tokenization: The chat is split into chunks of ~15,000 tokens each
2. Rolling Summarization:
   • First chunk is summarized using the seed prompt as context
   • Each subsequent chunk is summarized using the previous summary as context
   • This maintains narrative continuity throughout the conversation
3. Final Output: The last summary contains the condensed version of the entire conversation

The seed prompt (seed_prompt.txt) provides initial context about the conversation. It should include:

• Key participants and their roles
• Overall situation/context
• Important background information

The summary prompt (default: default_summary_prompt.txt) instructs the LLM on how to summarize. It can be customized to focus on specific aspects like:

• Emotional progression
• Key decisions
• Action items
• Relationship dynamics

# Export your chat from ChatGPT/Claude/etc as markdown
# Create a seed prompt with context
echo "This conversation is between a software developer and an AI assistant..." > seed_prompt.txt

# Run the summarizer
uv run python main.py my_chat_export.md --seed-prompt seed_prompt.txt --debug

# Review the final summary
cat final_summary.txt

• Uses GPT-4o for summarization
• Token counting via tiktoken
• Chunk size: 15,000 tokens (configurable in constants.py)
• Output length: ~1000-1200 words per summary

llm-chat-summarizer/
├── main.py              # Main orchestration
├── constants.py         # Configuration constants
├── file_io.py          # File reading/writing operations
├── tokenizer.py        # Text chunking and tokenization
├── summarizer.py       # OpenAI API calls and summarization logic
├── default_summary_prompt.txt  # Default summarization instructions
└── README.md

MIT