Corbenic AI puts exact model memory on local SSD

Belgian startup Corbenic AI says its Galahad software can make an AI server reuse work it has already completed instead of sending the same document through a model again. The system stores a model’s key-value cache on encrypted local NVMe storage and reloads it when the exact same input returns. ¹²

Galahad entered beta on October 1, 2026. Corbenic describes it as a per-server software layer for existing GPU machines, rather than a new model or a cloud service. Its two main components are Taliesin, which restores the cached model state, and Merlin, which identifies reusable content. ¹²

How SSD-backed AI memory works

When a language model reads a prompt, it builds a key-value cache that represents the attention state created during that work. Engines can reuse that state while it remains in GPU memory, but the cache disappears when memory pressure evicts it or when the serving process restarts. ¹³

Corbenic’s approach fingerprints the exact input bytes, model and tenant, then stores the resulting state locally. When the same material appears again, Galahad checks the fingerprint before loading the saved state. Its fail-closed rule is straightforward: if the check does not match, the system recomputes the answer rather than risk returning the wrong memory. ¹³

The company says no special hardware is needed beyond a Linux server, an Nvidia GPU and a standard NVMe SSD. It also says customer data stays on the customer’s infrastructure. Corbenic’s official product material describes Taliesin as a bit-exact restore layer and Merlin as a deduplication and lookup system. ¹²

The 14x claim needs careful reading

The eye-catching benchmark is a claimed 14x speed-up, but the source paper gives that figure more context. The result came from a recall test involving 100 facts hidden in a 96,726-token list, using Gemma 4 31B. Galahad combined Taliesin with Blaise, a retrieval component that passed the model roughly 668 relevant tokens rather than the whole corpus. ¹³

The paper credits Blaise with most of the gain. Taliesin alone was not faster in every comparison: on vLLM, its test took 15.56 seconds per question against an 8.11-second baseline, while llama.cpp improved from 9.25 seconds to 3.01 seconds in the relevant test. Those figures also compared different answer coverage, so they are not a clean universal measure of SSD caching. ¹

That puts Galahad in a crowded field. LMCache already supports moving key-value caches into CPU memory, local SSDs and remote storage, while other commercial and open-source projects are also targeting storage as a way to reduce AI inference costs. ¹⁴

Security is the sharper pitch

Corbenic’s more distinctive argument is exactness and isolation. The company says each tenant’s data is separated, saved state is encrypted and a failed verification is treated as a cache miss. That matters because previous research has shown that shared key-value caches can expose information through timing side channels in multi-tenant serving systems. ¹⁵

The caveat is that Corbenic’s headline benchmarks were produced by the company’s founder on rented cloud GPUs, and the work is presented in self-published papers rather than peer-reviewed research. Raw logs were not yet public in the report reviewed here. A proper independent comparison with LMCache, using identical hardware and each system’s normal prefix caching enabled, is still needed.

Our opinion

Galahad is an interesting answer to a real waste problem: paying a GPU to reread the same contracts, manuals or code every time a process restarts. But the 14x figure should not be treated as a general-purpose promise. The most convincing part of Corbenic’s pitch is its conservative failure mode and its focus on exact reuse; the least proven part is how much advantage remains once competing cache and retrieval systems are tested on the same workloads. Until those comparisons arrive, Galahad looks less like a revolution in AI memory than a carefully engineered storage layer with a useful security story.

For operators, the practical question is not whether SSDs can replace GPU memory. They cannot. It is whether verified local persistence can avoid repeating expensive work after a restart without compromising correctness. Corbenic has made that case plausible, but it still needs independent evidence.