Liquid Memory

A drop-in proxy that compresses your AI prompts on your own machines, cutting your bill across OpenAI, Anthropic, and Gemini.

Liquid Memory sits between your code and the model. You change one line, the base URL your app sends to, keep your key and your model, and every prompt is compressed on your own machines before it leaves your network. The text it keeps is byte-exact, never paraphrased, so what the model reads is not silently altered.

What it does

Compresses long prompts so you are billed for far fewer tokens.
Runs on your machines. Your prompts and provider keys never reach Liquid Memory.
On SWE-bench Verified, agents through Liquid Memory resolve the same tasks as going direct, with no quality loss on that benchmark.
Never costs more than calling the model directly: if a call would not save enough, your full prompt is sent unchanged.
Optional routing to a cheaper model, in-path guardrails, a tamper-evident audit log, and cross-session memory.

Honest limits

Compression is lossy at the document level but fact-preserving, and best on forgiving workloads (summaries, notes, long code context). It is not "lossless" and does not promise identical output on every prompt; the confidence gate sends the full prompt whenever compression is not safe.

Pricing

A commission only on the inference spend we provably remove. Nothing if we do not save you money.

Works with OpenAI, Anthropic, and Gemini. One line of code, no rewrites, no lock-in.