Invest · Kerneta

Memory management and shared processing.

Kerneta builds memory management and shared processing systems for AI. .dai is the first product: an open format that keeps an AI's memory in plain files you own. It is the foundation, not the destination. The same machinery that decides what an AI should remember also decides what work needs doing, what can be reused, and what never needed a model in the first place. We have several products to come in this domain.

We started from a frustration that is a choice rather than a technical necessity: the most useful thing an AI can have is a memory of your work, and today that memory is lost when a conversation ends, locked inside one company's product, or affordable only if you can pay to re-read your entire history on every question. We are raising our first round.

What accessible means to us

Four things, and we hold ourselves to all of them

01

Affordable enough that cost stops being the reason you stop

A developer working with an AI hits the context wall within weeks, and pays more for every question as the project grows. Recalling from .dai instead of re-reading everything means a question reads a mean of 10,065 tokens of context rather than a 103,601-token history. That is measured, with the arithmetic shown.

02

Yours, in files you can open

Your memory is plain text on your disk. Read it in any editor, grep it, put it in git, delete it. No database to run, no server required, no company standing between you and the record of your own work. If we disappeared tomorrow your files would still open.

03

Not tied to one AI company

The same store works with Claude, GPT, or a model running on your laptop. The format is portable; the accuracy is not identical. The same store spans 14.0 points across the five answering models in our cross-actor run, and we publish that spread rather than the best row in it. Memory portability is becoming the thing the big labs compete on; we think it should belong to you rather than be a feature one of them grants you.

04

Where it does not help

Our software will tell you not to convert a document when pasting it would work better, because that is what the measurements say. A memory system that cannot name its own limits will happily bill you inside them. Every number we publish traces to a run you can reproduce, including the ones where we lose.

The people

Who is building this

A small team, working in the open, publishing the evidence as it goes.

AR
Amin Rigi
Founder

Amin's doctoral research at the University of Edinburgh was on biosensing and wireless, battery-free RFID sensing systems: building instruments that measure something real without a power source of their own.

In 2012 he built the world's first lifeguard robot. His research and products have won 15 national and international awards. He is also co-founder of Eyesight Electronics, which treats amblyopia (lazy eye) and other eye conditions, and reports faster recovery at lower cost than the treatments it replaces.

The thread through all of it is the same: saving and improving lives with technology, and making it cheap enough that the people who need it can actually have it. That is the core of this work too.

At Kerneta he created the .dai format and built the retrieval engine behind it, taking it from a prototype to a measured 83.00% on the official LongMemEval benchmark with GPT-4o answering, using plain text files instead of a vector database: 22.40 points ahead of the full-context baseline on the same model. He set the rule the project runs on, that every claim traces to a run anyone can reproduce, and the results that go against us are published beside the ones that flatter us.

AM
Ali Munir
Co-founder

Ali specialises in bringing artificial intelligence and machine learning into the real world through embedded systems engineering. With a Master's degree in AI and work spanning AI/ML, real-time systems and IoT architecture, he has worked across the full stack: from training and deploying machine learning models to implementing them on resource-constrained hardware that actually has to run in production.

His AI work has ranged from environmental sensing to designing systems where intelligence runs at the edge, on microcontrollers and IoT devices with minimal power and memory. He also led embedded systems engineering across multiple hardware platforms while maintaining the compliance required for production work, communication protocols, and the kind of debugging and troubleshooting that separates working prototypes from shipping products.

The core challenge he solves is the same one that defines modern AI systems: making sophisticated computation work in constrained environments, limited power, limited memory, real-time requirements, without compromising reliability. That is where AI meets engineering: not in the theoretical space, but in the constraints of the physical world.

Investment

We will be raising our first round shortly

Kerneta is preparing its first external funding round. If you invest in developer tools, open formats or AI infrastructure and would like to look at what we are building, we would like to hear from you early, while there is still room to shape it.

We need capital that can be deployed quickly. This field moves in weeks, not quarters: the measurements that make our case are current now, and the work they fund (the hosted product, the parsers, the team features) is ready to start. We are therefore prioritising investors who can move fast.

Who we are looking for

We are looking for like-minded people who want to see AI accessible to everyone, not only to those who can afford it. That is the whole reason the format is plain text, the engine is open, and the cost of a memory is pennies rather than a subscription.

We also want to be plain about what we will not do, before anyone spends time on us: no military applications and no surveillance applications. Not as a current preference but as a condition. If that rules us out for a fund, say so at the first email rather than the last.

What we do want to put money into is strong R&D on applications that save and improve lives. That is the through-line of everything the founders have built before this, and it is what memory management and shared processing are for: making capable systems cheap enough to reach the people who need them.

Register your interest

Where we are

Built and measured so far

83.00%
on the official 500-question LongMemEval benchmark with GPT-4o answering, against 60.60% for full context on the same model
2nd of the reproducible
among memory systems whose configuration is reproducible, on LongMemEval-S with GPT-4o answering. Behind Mastra OM by 1.80 points, ahead of Supermemory. One closed system scores above us and we publish its number too (the full ranking, with every caveat)
3.0×
fewer tokens read per question than Mastra's published run: 10,065 against their published ~30k
Apache 2.0
the spec, the v4.4n engine and the MCP server, in one public repository, free permanently
$0.0009
to convert 1,000 tokens with the GPT-4.1-mini observer that produced every published number here (Opus 5 extracts comparably at about $0.0129 through the API, and free on a Claude subscription)

Every benchmark figure in this ledger links back to a measured run on the results page, with its method, its sample size and its caveats. Where a number is interpolated rather than measured, we say so on the page it appears. Figures in the founder biographies above come from that work, not from this project's runs, and are not covered by that guarantee.

Work with us

We are open to collaboration and partnership, and we want to work with people who will push the format further than we can on our own: teams building on .dai, organisations that want it inside their own products, researchers who want to test it against something harder, and anyone writing a reader in another language.

We are a small team, and a hard question helps us more than a compliment. If you think one of our numbers is wrong, tell us: that is the most useful message we can get.