How to Summarize Old Conversation Turns in Python

Compress old conversation turns into a brief summary while keeping recent turns intact for LLM context management.

Easy Python 3.9+ Aug 9, 2026 AI & LLM integration patterns 14 views 0 copies

llm context compression conversation summarization

Python code

43 lines

Python 3.9+

from datetime import datetime, timedelta


def summarize_old_turns(conversation, max_turns=5):
    """Compress turns older than max_turns into a brief summary."""
    if len(conversation) <= max_turns:
        return conversation, ""

    old_turns = conversation[:-max_turns]
    recent_turns = conversation[-max_turns:]

    # Build a compact summary of old turns
    speaker_counts = {}
    for speaker, text in old_turns:
        speaker_counts[speaker] = speaker_counts.get(speaker, 0) + 1

    summary = "Earlier: " + ", ".join(
        f"{speaker} spoke {count} time{'s' if count > 1 else ''}"
        for speaker, count in speaker_counts.items()
    ) + "."

    return recent_turns, summary


if __name__ == "__main__":
    mock_conversation = [
        ("alice", "Hello!"),
        ("bob", "Hi alice"),
        ("alice", "How are you?"),
        ("bob", "Fine, thanks"),
        ("alice", "Great to hear"),
        ("bob", "What about the project?"),
        ("alice", "It's on track"),
        ("bob", "Perfect"),
    ]

    recent, summary = summarize_old_turns(mock_conversation, max_turns=3)

    print("Summary of old turns:")
    print(summary)
    print("\nKept recent turns:")
    for speaker, text in recent:
        print(f"{speaker}: {text}")

Output

stdout

Summary of old turns:
Earlier: alice spoke 3 times, bob spoke 2 times.

Kept recent turns:
alice: It's on track
bob: Perfect

How it works

The function separates old turns from recent turns using list slicing. It counts how many times each speaker contributed to the old turns and builds a compact summary string. This pattern mimics token-efficient context compression for LLM conversations. The function returns both the recent turns and the summary, which you can prepend to your prompt or messages array.

Common mistakes

Forgetting that slicing keeps the original list order — old turns must come before recent turns
Using the same list variable after slicing and accidentally mutating the original
Assuming the summary length is always bounded — speaker counts can grow with many speakers

Variations

Use a summarizer LLM call to generate a natural-language summary instead of the count-based heuristic
Store summaries as a rolling buffer so you only summarize once instead of repeatedly

Real-world use cases

Trimming long chat histories before sending them to an LLM API to stay within token limits and reduce cost.
Maintaining a rolling context window in a customer-support bot while preserving key facts from earlier messages.
Building multi-turn agent memory systems where older exchanges get condensed into a short state snapshot.

How to Summarize Old Conversation Turns in Python

Python code

Output

How it works

Common mistakes

Variations

Real-world use cases

More from AI & LLM integration patterns

Tutorials

Quizzes

Python code

Output

How it works

Common mistakes

Variations

Real-world use cases

More from AI & LLM integration patterns

Keep learning

Tutorials

Quizzes