Comparison guide
SuperCompress vs sliding window
Sliding windows keep only the most recent N tokens of conversation history. That means old-but-important context gets dropped. SuperCompress scores every line against the current question instead, keeping relevant history regardless of how many turns ago it appeared.
How sliding windows work
A sliding window keeps the last N tokens of conversation history, discarding everything older. If a user asked about their account settings 20 turns ago and then asks a new question about billing, the sliding window has already dropped the account context.
With SuperCompress, every line is scored against the latest query. Old context that remains relevant — like the user's account type, previous issues, or preferences — is preserved. Irrelevant recent chatter is dropped instead.
Benchmark comparison
| Method | Oracle Recall | Query-Aware | Latency |
|---|---|---|---|
| Sliding window (4K) | ~35% | No | 0ms |
| Sliding window (8K) | ~45% | No | 0ms |
| SuperCompress | 100% | Yes | ~60ms |
When to use each
Sliding windows are appropriate when conversation recency is the only relevance criterion — for example, a simple Q&A bot where each turn is independent. SuperCompress is better when historical context matters: multi-turn agent interactions, customer support conversations, and complex problem-solving sessions.
For best results, combine both: compress first with SuperCompress, then apply a generous sliding window as a safety cap.
Frequently asked questions
Can I use SuperCompress and a sliding window together?
Yes. Compress first to remove irrelevant content, then cap the result with a sliding window at your maximum context budget. This gives you the best of both approaches.
Does a sliding window lose information permanently?
Yes. Sliding window truncation drops the oldest tokens. Once dropped, that information cannot be recovered for future turns.
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.