Common questions about Compresh, how it works, and what to expect.
What is Compresh?
Compresh is context-compression and episodic-memory middleware for LLM APIs. It sits between your application and your LLM provider, reconstructing prior context instead of resending the whole conversation each turn. One line to integrate — change your base URL, keep everything else the same.
How does compression work?
As a conversation deepens, Compresh reconstructs what mattered — decisions, facts, corrections — instead of retransmitting the full history every turn. The classification that guides this is an internal process: the model receives reconstructed context, never the tags. Short, single-turn requests pass through with minimal processing; deeper conversations see far more savings.
Will it affect response quality?
In our benchmark — 360 real-world Q&A items replayed as one long, growing session — Compresh cut input tokens by ~66% with no measurable quality loss versus sending the full history, and the memory layer added no token overhead. Short conversations (a handful of turns) see minimal compression — Compresh avoids compressing when there isn't enough context to do so safely. Deeper episodic recall is under active benchmarking; we publish only what we've measured.
Is my data safe?
Yes. API keys are encrypted with Fernet (SHA-256 derived key). No conversation content is stored permanently — Compresh processes context in-flight, and the internal compression state is transient (auto-expiring). Compresh never logs message content.
Can I use free or local models?
Yes. Free models and local LLMs (Ollama, LM Studio, etc.) work through Compresh. When you use a free or local model, there is no savings-share deduction — Compresh stays out of the way. The Starter tier requires a $10 minimum top-up to start (paid at $7.50 with the 25% discount), and a $5 minimum balance must remain on the account to keep free-model access active.
What's the pricing?
Starter: $0 service fee. $10 minimum top-up to start ($7.50 with the permanent 25% top-up discount), no card on file. Every verified account gets $30 in credit. Accounts using the MCP server or OpenClaw hook also include a 5-day free TUL 2.0 trial — no card, no commitment. Paid models on Starter: 30% savings-share. Pro: Subscription plans — Quarterly $18 (20% share), Semi-annual $33 (16% share), Annual $60 (12% share). Pro pays itself off if you use paid models heavily. See Pricing for full details.
What if Compresh doesn't recognize my model as free or paid?
Compresh detects model pricing through provider responses. Local model usage is auto-detected (no provider call goes out). For new or rare models we don't have pricing data on yet, we fall back to OpenRouter's lowest published tariff as a conservative default. If you believe a model is misclassified, you can dispute via support — we run a quick check and adjust.
Can I self-host?
The core proxy is open-source. You can fork the repository, configure your own keys, and deploy anywhere. Self-hosting means you handle infrastructure, updates, and scaling yourself.
Which models are supported?
Any OpenAI-compatible model, all Anthropic Claude models, and 200+ models via OpenRouter. If your provider exposes an OpenAI-compatible API, it works with Compresh. See Integrations for setup guides.
What about injection attacks?
Compresh includes a 3-layer injection detection system: regex pattern matching, heuristic analysis, and ML-based classification. It supports 19 languages out of the box. Injection detection runs on every request before it's forwarded to your provider, adding a security layer that most direct API integrations lack.