Galahad AI memory SSD claims 14x speed when it reads far less text

Galahad saves a model’s KV cache to encrypted local storage and reloads it when the same text shows up again. The KV (key-value) cache is the attention state a model builds while it reads a prompt, a kind of scratchpad of what it has already understood.

Normally, when a model gets a second question about the same document, it rebuilds that scratchpad from the first word. That step is called prefill. Serving engines like vLLM and SGLang already keep the scratchpad in GPU memory for reuse, but it gets thrown out when memory runs tight, and it vanishes completely if the server restarts.

Galahad’s storage piece, called Taliesin, writes each block of that state to disk. Every block is labeled with a fingerprint of the exact input bytes, the model and the customer, and is pulled back in only when all three match. Founder and CEO Sietse Schelpe frames the trade-off plainly: the real choice, he says, is between the SSD and redoing all the work on the GPU, and the load happens once, before the first word of the reply.

Corbenic says it needs nothing exotic. A Linux machine, an Nvidia GPU and an ordinary NVMe drive are enough. Schelpe also admits two limits up front: short prompts aren’t saved because recomputing them is already quick, and network storage is far slower than a drive inside the server. In the paper’s H100 test, a RAM disk and a local disk gave almost the same throughput, which is why he says faster storage didn’t make Galahad faster.

The 14x number comes from a test that combined Taliesin with a second tool, Blaise, rather than from SSD loading on its own. Schelpe points to 0.59 seconds per question on vLLM as the headline result.

That figure came from a recall test with 100 facts hidden inside a 96,726-token list, run on Google’s Gemma 4 31B model. Blaise keeps documents as plain text and passes the model only the slice a question needs, about 668 tokens instead of the whole pile. Corbenic’s paper itself credits Blaise with lifting the llama.cpp gain from 3.1x to 14.5x.

Take Blaise away and the picture flips for two of the three engines:

• vLLM: 15.56 seconds per question with Taliesin alone, against 8.11 seconds without it.
• SGLang: 22.02 seconds with Taliesin, against 10.29 seconds.
• llama.cpp: the only one that sped up, at 3.01 seconds against 9.25.

Even those comparisons aren’t like for like. The baseline ran with a 12,000-token window that could see only the last 12% of the list, and it got 10 of 100 questions right. With Taliesin, the model could see everything and got 98 of 100. So the slower runs were also far more accurate. That’s a fair trade in many jobs, but it isn’t the same as a clean speed test.

Parking KV cache on storage is already a busy field, and Galahad is arriving late. LMCache, an open-source project under the Apache 2.0 license with around 10,900 GitHub stars, already shifts KV cache into CPU memory, local SSDs and remote storage. Back in 2024 its blog pitched the idea as letting models read each text only once, close to Corbenic’s own slogan.

Samsung engineers described a restart-proof NVMe tier for LMCache just over a week before Galahad’s launch. Tensormesh, the company behind LMCache, raised $20 million in May from backers that include AMD Ventures, CoreWeave and Nvidia’s venture arm.

Nvidia’s Dynamo framework layers KV cache across GPU memory, system memory, local NVMe and networked storage, and the chipmaker showed a dedicated context memory platform, now called CMX, at CES 2026. WEKA, Phison and Huawei all sell their own versions of the same idea.

When asked why nobody had built this before, Schelpe said rivals had chased bigger models. The list above suggests plenty of people did build it. What Galahad claims instead is exactness: a saved block is reused only if it would give bit-for-bit identical output, and anything that fails a check is simply recomputed. The company sums up the rule as serving without memory rather than with wrong memory.

Corbenic’s launch materials actually lead with security rather than speed, describing Galahad as a locked vault for a model’s working memory. Stored state is encrypted, each customer’s data is kept apart, and developers can roll a server’s saved memory back to an earlier point.

That matters because researchers have shown that shared KV caches in popular engines can leak. By timing how fast answers come back, one user could piece together another user’s prompts. Galahad won’t start without a customer identity and a model fingerprint, and Schelpe says nothing leaves the server or goes back to Corbenic.

There are reasons for caution, though. Every number in the main paper comes from tests the founder ran himself on rented cloud GPUs, as the paper discloses. Raw logs haven’t been published yet, the method Blaise uses to pick text is being held back while patents are pending, and the papers are self-published preprints rather than peer-reviewed research.

You won’t install Galahad, but tools like it decide how quickly a company’s chatbot answers a follow-up question about a long PDF and how many GPUs that company has to rent to do it. If saved memory works safely, services can keep long documents warm between questions instead of paying to reread them.

The more useful lesson for everyday users is about retrieval. Schelpe compares Blaise with search-based systems that hand the model only the top-ranked chunks of a document. If that search misses, the model never sees the answer. In his cited benchmark, Galahad answered 100 of 100 questions against 77 for the RAGFlow pipeline. That gap, if it holds up, is the difference between a work assistant that finds the clause you asked about and one that confidently says it isn’t there.

The next real test is an outside one: a head-to-head with LMCache on the same Nvidia hardware, with each engine’s own caching switched on, and Blaise tried on documents it wasn’t tuned for.