One agent, many workspaces: where shared memory fits
An agent can work across a business without carrying every project's assumptions into every task. Here is where workspaces, skills, and Space each help.
The hidden cost of an agent for every function
A founder asks an AI agent to help with marketing. Soon there is a sales agent, an SEO agent, an engineering agent, and another agent to coordinate them. Each has its own instructions and some memory of the business. Keeping all of them current becomes work of its own.
A permanent agent for each function can also carry the wrong context between projects. Imagine one marketing agent serving two startups. In one, enterprise buyers respond to technical, premium positioning. In the other, the audience is consumers and the distribution strategy depends on community. The agent might reuse a successful pattern from the first project without saying where it came from.
The recommendation could sound sensible while resting on the wrong customer, offer, or evidence. The problem is not that the agent knows too much. It is that the system has not made the scope of that knowledge clear.
Where Space fits
Local instructions tell an agent how to work. They do not hold every decision a project has made. Chat history holds some of those decisions, but finding the right message later can be unreliable, especially when work continues in a new session or another tool.
Space is the durable context layer in this design. An authorized agent can use its MCP connection to work with documents, tasks, projects, and memory. A decision can be saved with the work it concerns, then retrieved in a later session. A person can inspect and correct saved memory rather than treating an agent's recollection as unquestionable fact.
Space is external to the agent's persona. That matters when an operator changes models, delegates a task, or resumes work after the original conversation is gone. The context belongs to the workspace, not to one chat transcript.
One agent can work in distinct project workspaces
The agent must identify the current project and check whether a retrieved claim applies. Project organization aids relevance; the selected Space and its grants control access.
Give memory a scope too
A founder's preference for concise reports might apply across their work. A customer's objection to annual contracts belongs to a particular project. The result of a landing page experiment belongs to that experiment. Saving all three as undifferentiated global memory would make later retrieval harder to judge.
Space can connect work to projects and organize memory with projects, categories, and sections. It can also retain source information for claims derived from saved documents, so a reader can inspect what an earlier revision actually said. A remembered sentence may still be incomplete, outdated, or misclassified. Its source and scope help the agent and person recognize that.
Project organization is a relevance aid, not a permission boundary. The selected Space and its grants define access. If two businesses require separate access, they should be treated as separate access scopes rather than relying on project labels alone. Within an authorized Space, the agent should still name the project in its search and check the scope of the results before using them.
Sources: Organize memories with projects and categories; Inspect memory sources and corrections
Let projects learn from each other deliberately
Suppose a positioning experiment in Project A produces an idea that might help Project B. A main agent that sees both projects can notice the connection. It should then compare the audiences, constraints, and evidence before proposing the idea for B.
That creates a deliberate transfer: this worked there, here is why it might or might not work here, and here is how we could test it. Space helps keep the original result attached to Project A while giving the agent a place to record a new hypothesis or decision for Project B.
The aim is neither automatic sharing nor total isolation. It is to make the origin of a useful lesson visible when the lesson moves.
A lesson crosses projects only after a scope check
- SourceA result in Project AKeep the original experiment and its evidence attached to Project A.
- CompareCheck the analogyCompare audience, offer, channel, and constraints with Project B.
- TryForm a hypothesis for BAdapt the idea and test it instead of treating A's result as a fact about B.
- RecordSave B's outcomePut the new decision or result with Project B and link back to the original when useful.
Cross-project learning is a judgment the agent makes with evidence, not an automatic memory merge.
Bring in workers when the task calls for them
One main agent does not have to perform every step in sequence. It can ask temporary workers to research separate questions, implement bounded changes, or review a result. Hermes Agent, for example, supports delegated subagents and an independent review workflow.
A worker should receive the task's relevant project context and return its findings with enough evidence for the main agent to evaluate. The main agent remains responsible for integrating the work and saving durable outcomes. More agents do not automatically make a claim true; a second reasoning pass is valuable when it can inspect the source, run the test, or challenge the first answer.
A daily loop for an agent and its memory
The human interface could be Discord, a terminal, or another agent client. A founder might ask what is blocked, drop an idea into an inbox, or start a piece of engineering work. The useful sequence is the same: establish the current project, read local instructions, retrieve relevant records, do the work, verify the outcome, and save what should survive.
An inbox message does not have to become permanent memory. A confirmed decision, a changed task, or an unresolved blocker often should. Saving those outcomes gives the next session something more dependable than an unstructured scrollback.
This is an architecture we are exploring, not a claim that Space already runs an autonomous company workflow. Space is in development, and public sign-ups are not open. Our recorded Codex integration check covered authorized document and task operations in a development environment; continuity and retrieval across clients need their own checks.
One operator, with context that can outlast it
The goal is not to minimize the number of agents. Persistent separation makes sense for different permissions or independent responsibilities. Local instructions make sense for different projects and kinds of work. Temporary workers make sense when parallel execution or independent review improves a result.
Space addresses the continuity between those parts. It gives an authorized agent a place to find the work, a person a way to inspect what was saved, and the next session a way to pick up from a durable record. The architecture becomes easier to reason about when identity, methods, assumptions, and memory each have a clear home.