Questions, answered.

Yes — measured, not assumed. In a July 2026 head-to-head on a 1,000-document library (500 questions, identical model and retrieval on both sides), answer quality was statistically identical to standard RAG — a result we then re-checked with a second, independent judge (not the model grading its own work), which returned the same dead tie. On single documents, answering from the saved memory matched the quality of re-reading the entire document, where chunk-based RAG scored lower. Cheaper-but-worse isn't the product — we verify parity on your workload before it goes in front of your customers.

The changed document is simply read again — one forward pass — and its saved memory is replaced. No retraining, no index rebuild, nothing else is touched.

Open-weight LLMs you control (e.g. Qwen3, Llama), served through a pip-installable plug-in for vLLM — the standard open-model serving engine. The model stays frozen; document memories are saved KV, not weight changes. Nothing proprietary enters your serving path, and no fork of your serving stack.

Inside your own cloud account, with no public ingress. Your customers' documents never train a shared model and never leave your perimeter — a data-residency guarantee you can pass straight through to them.

Fine-tuning bakes facts into weights — training runs, hard to update, prone to hallucination, no sources. Engram Smart CAG involves no training at all: knowledge lives in each document's saved memory, swappable per document in seconds, and every answer is grounded in a specific source.

Sub-second to single-digit seconds per document — one forward pass through the model, no training. A 10,000-document library onboards for roughly $9 on spot GPUs, and the per-answer savings repay that within normal usage.

Think of it as your prefix cache made durable and addressable. The built-in cache is ephemeral — keyed by exact token prefix, evicted whenever the pool churns, gone on restart, and the client must re-send the full document text to rebuild it. Engram's document memories persist in tiered storage, page onto any instance in the fleet, and are addressed by document id — a request of a few dozen tokens instead of re-shipping thousands.

Saved memories change size with the model you're running. Our own compression codec shrinks them 4×, with answer quality statistically unchanged. By adding automatic cloud tiering, costs can be mitigated to nearly match a RAG index's storage costs, so the gap has essentially closed. Storage costs are quickly offset by the GPU savings.