Galahad-KV

Long-term memory for AI models — a KV-cache layer for vLLM and SGLang.

Galahad keeps the KV cache your model already computed, encrypted on disk, and gives it back on the next run — so a long prompt, a document, or a conversation is remembered across requests and across restarts instead of being recomputed every time. It installs with one command and the serving engine finds it automatically.

pip install galahad-kv

Linux x86-64, Python 3.10–3.14. libgalahad.so ships inside the wheel — no compiler, no build step.

Beta. Galahad-KV is launching as a beta with paid pilots. It is in active use and tested on vLLM and SGLang; limits are stated plainly and are being improved through pilots.

Quick start

Every install needs a licence — it is also the encryption key for your stored memory, and it never leaves your machine.

# 1. a free licence (one GPU, non-commercial, 12 months) — or `galahad buy` for more
galahad free --org your-org-key --email you@company.com --accept-non-commercial

# 2. check everything is ready
galahad doctor

Then activate it in whichever way you serve:

vLLM

export GALAHAD_CACHE_DIR=/var/galahad
vllm serve <model> \
  --kv-transfer-config '{"kv_connector":"GalahadConnector","kv_role":"kv_both"}'

vLLM discovers the connector by itself — no module path, no PYTHONPATH.

SGLang

export GALAHAD_CACHE_DIR=/var/galahad
python -m sglang.launch_server --model-path <model> \
  --enable-hierarchical-cache --hicache-ratio 1.5 --hicache-storage-backend dynamic \
  --hicache-storage-backend-extra-config \
  '{"backend_name":"galahad","module_path":"galahad_vllm.sglang_backend","class_name":"GalahadHiCacheStorage"}'

Your own code (bare)

Link libgalahad.so and call its C ABI directly; the licence supplies the key. See the wiki for the five-minute check.

The galahad command

pip install galahad-kv gives you one command for everything:

What you get

• Memory that survives restarts. What was remembered on one run is available on the next, without recomputing it.
• Encrypted at rest. Your stored memory is encrypted. By default your licence unlocks it; nothing extra to manage.
• Your key stays yours. Optionally hold your own key via GALAHAD_STORE_PASSPHRASE — it stays on your machine and is never sent to us. Lose it, lose the memory; that is the point.
• One install, both engines. The same wheel works with vLLM and SGLang.
• No surprises without a licence. Without one, it will not quietly do nothing — it tells you.

Known limits (beta, stated plainly)

• One serving process = one trust domain. Do not put mutually-untrusted users behind one server; run separate processes, stores and keys per customer.
• Version pins. vLLM 0.30, SGLang 0.5.20, matching transformers. Newer engine versions change internal APIs the connector binds to.
• Linux only (manylinux_2_28, glibc ≥ 2.27), Python 3.10–3.14.

Links

• Documentation & guides: the project wiki
• Issues & contact: see the project page

© 2026 Corbenic AI — Sietse Schelpe. All rights reserved. Galahad-KV is proprietary software; see the licence terms delivered with your key. Author: Sietse Schelpe — Corbenic AI.