Why it exists
The working experience that produced this project is specific and probably familiar. You spend months on a codebase with an assistant that is genuinely good at the work. It is also, every single morning, new. It does not know that you settled the API shape in March, that the migration was reverted for a reason, or that this team squashes before merging because of a decision taken after a bad week in April.
So you re-explain. The re-explaining is not the expensive part — a paragraph of context is cheap. The expensive part is that you stop trusting the collaboration with anything that depends on continuity, and you route around it, and what could have been a colleague becomes a very fast intern who is permanently on their first day.
The instinct is to fix this in the model: a longer context window, a bigger system prompt, a file the model is asked to read first. Those help and they do not solve it, because they all put the burden of continuity on the thing that resets. Continuity belongs to a layer that persists.
That is the whole thesis. The model brings intelligence; the layer underneath brings practice. Keep them separate and each one can be good at its job.
The name is literal
Myelin is a fatty sheath that wraps a nerve fibre, laid down through repetition, and it is why practiced skills run faster than new ones.
A bare axon conducts by propagating the signal continuously along its whole membrane. A myelinated axon is wrapped, with small gaps between the wraps — nodes of Ranvier — and the signal jumps from node to node instead of travelling every micron in between. This is called saltatory conduction, from saltare, to leap. Signal speed jumps roughly a hundredfold.
The important part is how it gets faster. Not more force, not a better signal: the pathway stops having to traverse what is already established. That is precisely the thing this product does, and it is why the name is a description rather than a metaphor.
Practice does not change what you are capable of thinking. It changes how much of the path you still have to walk.
What it is not
Stated because the category is crowded and mostly with different things.
Not an agent. Myelin does not act, plan, or decide. It answers questions about what this brain holds and refuses writes that break its rules.
Not a framework. There is nothing to build on top of and no application structure to adopt. It is an MCP server; your host talks to it.
Not a model. It never calls one, and it holds no model API keys because there is nothing to call. Every behaviour is deterministic: storage, retrieval, ranking, validation, refusal.
Not a vector database with a wrapper. Retrieval today is FTS5 plus trust ranking plus one hop of edge expansion. That is fast, explainable, and free.
The design commitments
Four decisions the rest of the system follows from, kept in one place so they can be argued with:
Determinism over cleverness. Anything that could be a model call and does not have to be, is not one. The cost is that Myelin is less flexible than a system that could reason about your intent. The benefit is that it gives the same answer twice, runs in milliseconds, and can always tell you which rule fired.
Append-only over updates in place. History is the part you need later, and it is the part an update destroys.
Proposals over autonomy. The system can notice a pattern; it cannot act on one. Anything that would change what Myelin does in future sessions requires a person.
Say the caveat in the same breath as the claim. The security page is the clearest case: the encryption story is genuinely strong and there are two ways it does not save you, and both are on the page in plain language rather than in a footnote.
Status
Pre-release, under active development.
There is a hosted instance at myelin.myelin-layer.com. It is one small server
run by the project owner, access is invite-only because there is no self-serve
signup, and it carries no uptime guarantee and no support commitment. There is
no free tier, no paid tier, no waitlist, and no company, because there is no
business here — there is a server and a person who runs it.
You can also ignore it entirely. Running over stdio on your own machine needs no account and no network, and running your own server is documented end to end. If a future version of this page describes something else, it will say so.
What exists today: the five cores, both transports, the eighteen MCP tools, the CLI, the encryption and key handling described on the security page, and a deployment runbook that has been followed end to end.
Roadmap
Short and honest, which means it lists one thing rather than ten.
Vector search is v2. Today’s recall is FTS5 plus trust ranking, which is fast, explainable, and requires no model. Semantic retrieval is a real improvement for a real class of query, and it brings an embedding model, an index to keep in sync, and a scoring function that is much harder to explain when it returns something odd. It goes in when it earns that, not before.
Everything else that might be listed here would be speculation, so it is not listed.
The project
Written in Go. MIT licensed. The source, the build log with the reasoning behind each decision, and the deployment runbook are all in the repository.
This site collects nothing. No analytics, no cookies, no third-party requests, no fonts loaded from anyone else’s server. Reading it is not an event that gets recorded anywhere.