A format is only open
if other people use it.
We can publish a specification and an engine. We cannot make .dai a standard on our own. That takes other implementations, other readers, other people checking our numbers and telling us where we are wrong.
Come and build on it. The format is open, the engine is Apache 2.0, and the useful work now is everything that connects .dai to the rest of your world: apps, connectors, plugins, editors, agents, whatever you already use. Tell us what you need and we will build the ones that matter most, together with you. No membership and no newsletter you have to escape from later: The hosted converter has a sign-in, because metering and billing need one, but nothing here does, and neither does the format. You should not need an account to own your own memory, and that is still the argument.
Ten ways to take part, from ten seconds to a weekend
The first three need no code at all, and they are the ones that help most right now. None of them need an account with us. Several of them route through GitHub, and the repository is public, so those are open now. The email below works just as well if you would rather not use GitHub.
Tell us what you need built
The fastest way to shape this. Name the app, the connector, the plugin or the workflow that would make .dai useful in your day, and we will build the ones that matter most, most-asked first. You do not need to write any of it yourself: saying what is missing is the contribution.
a minute · email or GitHub DiscussionsConnect it to something we have not
Anything that speaks MCP already works. Everything else is open and wanted: ChatGPT, an editor, a framework, a CLI, a mobile app, a plugin for the notes app you already live in, a cloud surface. Build one and we will help, link it, and keep it working as the format moves.
a weekend · anything you already useShare it, and star the repository
The least glamorous and the most useful. A star is how other people find this at all, and one link from someone who actually uses it is worth more than anything we can say about ourselves. Send it to the person you know who keeps losing context.
ten seconds · GitHub, or anywhere you talkTell us it broke
The most useful report we get is "I tried to do X and this got in the way". Not a feature request in the abstract, a real thing that failed on your machine.
a minute · email or GitHub IssuesAdd yourself to the users list
This list is a file in the public repository, so you can open a pull request adding a line, which doubles as a first contribution if you have never opened a PR before. Or email us and we will add it for you. No account, no form.
five minutes · one line in a fileShow your store
Because a .dai file is readable text, you can paste one. What did your assistant actually remember, and did it get it right? This is the thing no vector database can do, and seeing real stores teaches us more about the format than any amount of internal testing.
five minutes · email it, or Show and tellAnswer someone else's question
Discussions will be public and indexed, so an answer you write once is found by everyone who hits the same problem later. That is why we will start there rather than in a chat server where answers disappear.
ongoing · Q&A, from launchWrite a reader in another language
The format is deliberately simple enough to implement in an afternoon: UTF-8, a YAML header, a JSON zone and text. A Python, Go or Rust reader does more for the format than almost anything we could ship ourselves.
an afternoon · your language of choiceMove a number on the benchmark
A change that improves accuracy gets published with the run that proves it, on the results page, whichever way it lands. The open areas, and what each one would need, are listed in the repository.
a weekend or more · the deep endBuild a benchmark we do not have
Several areas have no scoreboard at all: agent memory, document formats, small-window retrieval. Without a benchmark a contribution cannot be judged, so building one unblocks everyone working in that area, not just you.
a project · genuine researchEvery number here is a paid run
A single full 500-question evaluation costs about $17 in model spend to answer and to score with the benchmark authors' own judge, on top of roughly $46 to $59 to build the stores from a 64-million-token haystack. The stores are cached, so the runs after the first cost the $17. The version ladder on the results page and the actor comparison are the handful worth publishing. Behind them are hundreds of runs, re-runs and tests, and the ones that measured something that did not work cost exactly the same as the ones that worked.
That is what support pays for: more runs, on more benchmarks, against more models, published whichever way they land. Three ways to help, and the first one is free.
Star the repository
Stars are a vanity metric and they also genuinely work: GitHub ranks and recommends by them, so a star is the difference between the format being found by the next person and not. It costs nothing and takes about ten seconds. You need a GitHub account, which is free, and if you are benefiting from the format and have never made one, this is a good reason to.
It is the single highest-leverage thing here, because it spreads the format rather than funding us. A format nobody has heard of is not a standard.
ten seconds · free · needs a GitHub accountTake a supporter badge, and put your name in the file
Because this project's whole argument is that plain text outlives the software that wrote it, the badge is not a graphic on a dashboard. It is a line in SUPPORTERS.dai, a real .dai file in the repository, in the format itself. It lands in git history, it is greppable, and it is still readable when this website is gone. That is the point.
Sponsor the project
For companies using .dai in production, or anyone who wants to fund the work directly rather than take a badge. Sponsorship pays for benchmark runs, the independent evaluations we cannot currently afford, and the connectors people keep asking for. The form below registers interest and we come back to you personally with the details.
any amount · form belowListed supporter
Your name or handle, and a link if you want one, added to SUPPORTERS.dai in the repository and to this page. Permanent, in the git history, in the format itself.
- A line in
SUPPORTERS.dai, with your link - Listed on this page
- Named however you want, or anonymously
Founding supporter
Everything above, plus a say in the work itself. Limited to the people who back it before the format has proved itself, which is when backing it actually matters.
- Everything in the first badge
- Named in the release notes of the next spec version
- A vote on what we benchmark next, and on which connector gets built first
- The roadmap before it is public
Tell us below which badge you want and we will come back with the details. We will never add your name anywhere without asking you first.
Support the project
Badge request, sponsorship enquiry, or a question about either. We reply personally.
This records that you would like us to get in touch. It takes no payment and stores no card details. We use your details to reply to you and for nothing else, and we will confirm with you before your name appears anywhere.
GitHub Discussions, in public, from launch
GitHub Discussions is the venue. It will be public, searchable and permanent, which matters more than it sounds: an answer written in a chat server is gone in a week, and an answer written in Discussions is found by search a year later by someone with the same problem. Until then the email above is the way in.
On Discord, plainly
We will open one, later. A chat server with four people in it reads as a failure, so it waits until there are enough of us to fill it.
Until then the conversation happens in GitHub Discussions, where an answer is still searchable a year later instead of having scrolled away in a week. When there are enough users we will open a Discord and say so here, and asking for one is itself a useful signal.
The facts, ready to quote
Correct as of 6 August 2026. If you are writing about the project and want to check a figure, or want something that is not here, email community@kerneta.com.
| What it is | .dai, an open plain-text memory format for LLMs, AI assistants and agents |
| Built by | Kerneta |
| Licence | Apache 2.0: the spec, the v4.4n engine behind the figure below, and the MCP server, together in one public repository at github.com/Kerneta/daidocs |
| Benchmark result | 83.00% on LongMemEval-S (415/500), GPT-4o answering, engine v4.4n |
| Against the alternative | 60.60% for pasting the whole history into the window, GPT-4o answering |
| Context read per question | a mean of 10,065 tokens, against a mean 103,601-token history |
| Standing | second among memory systems whose configuration is reproducible |
| Scored by | the benchmark authors' own evaluate_qa.py, judge gpt-4o-2024-08-06 |
Two things we would ask you to keep attached to any number: the answering model (GPT-4o, for every figure above), because swapping it moves accuracy more than any memory system does, and the denominator (500 of 500, abstentions included), because scores on this benchmark are quoted on several different question counts.
Logos and assets
Logos, icons and the social card are available on request, in the format and resolution you need. Contact us and we will send them over. Please do not restyle the mark, place it on a background that removes its contrast, or imply we endorse something we have not seen.
Naming, so it reads correctly in print: the format is .dai, the product is DaiDocs, the company is Kerneta, and the site is daidocs.com. The engine is the Kerneta Engine, one implementation of the format rather than the format itself.
Writing about it? Email community@kerneta.com. We will help you check a number, explain a method, or find the run behind a claim.
Using .dai at work, or in something you built?
Tell us and we will list you here. A company, a side project, a research group, a tool that reads or writes the format: all of it counts, and a small weekend project is as interesting to us as a large deployment.
Say what you are building, how you are using the format, and how you would like to be credited. If you would rather not be named publicly, say so and we will keep it private. We will never add anyone without asking first, and we will remove you on request, no reason needed.
The list is empty because the project is new. It will fill up with real names or not at all.
Start anywhere
Read the code or install it in two minutes. Either door works, and nobody has ever been turned away for asking a question that turned out to be simple.