Advanced guide
Multi-turn context compression
Long conversations are the most expensive LLM use case. After 20 turns, the full history can exceed 10,000 tokens. Multi-turn compression keeps the signal from all turns while dropping the noise.
The multi-turn challenge
After 50 turns in a customer support conversation, the LLM prompt includes: greetings, pleasantries, status checks, internal notes, and off-topic discussions — all mixed with the actual problem-solving context. A customer might mention their account type on turn 3, then ask a billing question on turn 45. The turn-3 detail is critical but would be dropped by any recency-based approach.
Solution
from supercompress import Compressor
comp = Compressor()
def compress_conversation(messages, latest_query):
history = "\n".join(
f"{m['role']}: {m['content']}" for m in messages[:-1]
)
result = comp.compress(history, latest_query)
# Now use result.compressed_text as your conversation history
return result.compressed_textFrequently asked questions
Does this work for 100+ turn conversations?
Yes. SuperCompress handles conversations of any length efficiently.
Will it keep system prompts and instructions?
Yes. System-level messages are preserved when relevant to the current query.
Build with less context
Put compression in front of your next LLM call.
Use the hosted API or run SuperCompress locally. Keep the evidence, drop the token waste, and measure the savings before it reaches your model.