Compress the most expensive bytes in AI
An LLM’s KV-cache, activations and embeddings are float tensors with strong per-channel and temporal structure — and its context is mostly tokens the answer doesn’t need. AT-1 Memory compresses both, losslessly where it counts: KV/tensor bytes recovered byte-exact, and context pruned so the model’s greedy output stays byte-identical. Every container carries a SHA-256 you can re-verify.
- byte-exact
- lossless on real GPT-2 KV + activations, SHA-256 verified
- 1.24×
- real GPT-2 fp16 activations vs raw (1.15× beyond deflate)
- never-worse
- min-of-both container falls back on incompressible input
- output = ident.
- context mode: greedy continuation stays byte-identical
KV & tensor mode
A lossless, structure-aware codec for KV-cache, activation and embedding tensors. It reorders token-major floats to channel-major and splits each value into byte-planes, exposing the correlation that general compressors miss — then keeps whichever is smaller, so it is provably never-worse than plain deflate. Works on fp32/fp16/bf16.
at1 ctx kv compress kv_layer0.npy # -> kv_layer0.at1kv (structure-aware, byte-exact) at1 ctx kv verify kv_layer0.at1kv # SHA-256 checked
Context mode
Prune a prompt, system message, or RAG chunk so a given model’s greedy continuation is byte-identicalto the full context. Compress once, reuse across every query — the verification cost amortizes to zero. The .at1ctx container stores the kept token ids plus a SHA-256 of the guaranteed output, so anyone can re-verify the guarantee.
at1 ctx compress prompt.txt --model gpt2 # -> prompt.at1ctx (output-lossless) at1 ctx verify prompt.at1ctx # greedy output matches
What’s proven, and what isn’t
- Proven: byte-exact lossless roundtrip on real GPT-2 KV-cache and activation tensors at fp32 and fp16 (≈1.22–1.24× over raw, ≈1.13–1.15× beyond plain deflate), a never-worse fallback on incompressible input, and single-byte tamper refusal via the embedded SHA-256.
- Honest boundary:real transformer tensors are fairly high-entropy, so the lossless structure-aware win over deflate is real but modest (≈1.15×); the ratio scales with a tensor’s per-channel / temporal structure and is reported per tensor, never promised as a universal. The larger opportunity is the near-lossless (quantized) regime — on the roadmap, not claimed here.
- Not a confidentiality claim: AT-1 Memory is about size + integrity, not encryption. Pair with your own transport/at-rest encryption for secrecy.
Billed on value delivered — tokens saved (context) and byte-savings converted to token-equivalents (KV). First 1M tokens/month free. See pricing.