JournalCompany
One private memory for every agent you use
Builder is now free for one person and Teams is free for up to four. One private Spark memory, outside any model or application, that follows you across the agents you use and starts earning its place on the second task.
I don't use one AI agent for everything, and I doubt many serious users will.
One model is better at research. Another is cheaper for routine work. An IDE agent understands the code in front of it, while a custom workflow handles an operational task in the background. Six months from now, the mix will be different again.
The strange part is that every change of tool still means starting again.
Claude remembers Claude. Cursor remembers Cursor. A custom agent gets whatever prompt file someone remembered to attach. The same correction ends up copied into several places, each drifting away from the others.
That is backwards. Models, harnesses and IDEs should be replaceable. What you learn from using them should not be. Spark is model-agnostic, harness-agnostic and IDE-agnostic for exactly that reason.
We're changing the way people can start with Spark.
Builder is now free for one person, with a private memory you can connect to the agents you use. Teams is free for up to four people. The fifth person is where paid seats begin.
Your memory, not your agent's memory
A solo builder can easily be running half a dozen agents across different tools. That is already a team, even if all the human names on the account are yours.
When one agent learns that an API needs an idempotency key, another shouldn't have to rediscover it. When you correct the evidence standard in a research task, the lesson should be available to the next supported agent you ask. When you move from a frontier model to something smaller or local, you shouldn't lose the accumulated knowledge that made the first model useful.
Builder puts that memory in a private Spark space outside any one model or application. It works through MCP, so the memory can follow you across supported agents rather than being trapped in the tool that happened to create it.
Private means private by default. Your corrections do not go into a public commons. Memories keep their source and scope, and you can see what your agents are being told rather than relying on an invisible profile inside a model provider.
Four people is already a real team
The moment a second person arrives, memory stops being only about personal convenience.
Two people can give agents conflicting advice. A correction from one project can be dangerous in another. Nobody wants a private client rule leaking into unrelated work because a retrieval system decided the words looked similar.
That is why Teams includes private team memory across agents, tools and repositories, with the provenance and scoping needed to keep it usable. One person's useful correction can help the rest of the team, but it still has an owner, a source and a place where it applies.
Teams is free for up to four people. We chose that boundary because a small team should be able to prove the learning loop on its own work before procurement, a sales call or a budget line enters the picture.
Correct one recurring task, then run a different but related task later. Did the next agent use the lesson? Did the output improve? Did it get there without repeating the same investigation? If not, the memory count in a dashboard is meaningless.
Better economics without metering the learning
Spark doesn't charge per memory call, retrieval or agent run under normal use. Metering memory activity would be a dumb incentive: if memory is helping, your agents should use it constantly.
The saving appears in the task itself. An agent that can reuse a trusted lesson takes fewer unnecessary steps, finishes sooner and wastes fewer tokens. It also behaves more consistently because it is not improvising the same local rule from scratch on every run.
In our controlled ten-task SWE-bench evaluation, steady-state runs with Spark were 40% cheaper, 34% faster and used 31% fewer steps. Cost variance also fell by more than half.
Those are controlled coding results, not universal guarantees. They show the economic mechanism: once useful knowledge exists, the next agent can take a shorter path instead of paying to discover it again.
Start with the second task
Builder is free for one person. Teams is free for up to four. Both are designed around the same idea: one private memory for the agents doing your work.
Connect an agent, make a correction and then switch tools if you like. The next task is where Spark earns its place.
Context is rented. Memory is owned.