What is Vectadyne?

Vectadyne is not a vector database. Start there, because the category is the thing people get wrong first, and the mistake changes what you expect from it.

The problem

An engineering agent starts every run knowing nothing about your systems. It does not know that the payments migration was rolled back twice, that billing/reconcile.py is the file that breaks the nightly job, or that the team decided in March to stop using the legacy queue. So it rediscovers these things — expensively, in production, one incident at a time.

Retrieval-augmented generation gets you part way and then stops, for a reason that is structural rather than a tuning problem: similarity is not authority. A nearest-neighbour search over your documents will happily return a decision that was superseded last month, a lesson from a different repository, or a note whose author had no business writing it. It has no way to tell you which.

What Vectadyne does instead

Four things a vector store does not:

Types the memory. A decision is not a lesson is not a policy. Each kind has different rules about who may write it, how long it stays authoritative, and what happens when two disagree. A policy outranks a lesson; a decision can be superseded by a later one and stops being returned as current.

Scopes it. Every memory is pinned to a scope — an organisation, a repository, a service. A lesson learned in repo:payments does not leak into repo:marketing-site because they happen to embed near each other. Retrieval is admission-filtered at read time: what comes back is what this credential is allowed to see, and an empty result is a legitimate answer.

Keeps provenance. Every memory traces to the evidence that produced it — a PR, a CI run, an incident, a review. An agent can cite its reasons, and a human can audit them. A memory with no supporting evidence is visibly weaker than one with five.

Judges the action, not just the context. This is the part with no analogue in a retrieval system. Before an agent force-pushes, runs a migration, or modifies CI, it can describe the action and get a verdict — informed by what the corpus knows about that exact file, that exact service, and what happened the last three times.

Where it sits

The whole integration surface is an endpoint and a token. There is no agent to install, no sidecar, no daemon, and nothing to run inside your network. Vectadyne speaks MCP — so most agent frameworks connect to it with configuration rather than code — and the same capabilities are available over REST.

What it is not

• Not a vector database. It uses retrieval; it is not one, and swapping the backend does not change what it offers.
• Not a prompt manager or a context window packer. It answers what should be known here; how you spend your token budget is yours. It will measure that spend if you tell it.
• Not an inline enforcement layer. See Advisory-first.
• Not a model. It does not generate your code or your plan.

Next

• Quickstart — reachable in one unauthenticated call
• The adoption ladder — what you get at each rung, and what each costs
• What we do not claim — the limits, up front