Selective recall and memory scope
Two things decide what an agent knows: what was captured, and what is allowed to reach it. Both are narrower than people expect, and both are yours to set.
What reaches an agent on a recall?
The smallest relevant set of precedents, not the sessions that produced them. A briefing arrives as a few dense lines under a budget, so recall spends context rather than filling it.
Relevance is scored against what the agent is about to do. A lesson that has been superseded, retired, or quarantined stays out of the answer — unless the question is about history, in which case it comes back labelled with the window it was valid for. Recall and capture covers the mechanism.
Personal work stays personal
Your own memory is a separate graph from any team you belong to. Not a folder inside the team's and not a permission layered over it — a separate graph, with nothing linking across.
That separation is structural rather than a setting, which is what makes it worth trusting. A memory reaches a team when you promote it, and only then. When it is not obvious where a write belongs, it lands in your personal graph, because personal work can always be promoted later.
An agent working on your behalf can read both at once — your own experience and your team's — and every result says which one it came from. Writes go to one place; reads can span.
What a team shares
One graph across every agent, tool, and project on the team. Pooling is the point of a team plan: what the order desk learned in an incident is what keeps the warehouse agent from repeating it. Agent memory use cases walks through what that looks like by function.
A shared graph still does not mean everything reaching everyone. Memory can be grouped within a team, and each group granted to the people and agents who need it.
Narrowing what comes back
A recall can be pointed at less than everything:
- At one project. Ask against a single slice of a team's memory instead of the whole graph.
- At one source. Filter to a user, an agent, or a single session.
- At a smaller briefing. Ask for less when context is tight. A call can go under the ceiling, never over it.
Keeping work out of Causal
Capture is a call your agent makes, not a crawl. Causal sees a decision when your agent judges one worth keeping, and nothing besides — no repository scan, no background indexing, no upload of files it was never handed.
Where you install it is therefore the control that matters most: a project with Causal wired in builds memory, and a project without it writes nothing.
Removing something already remembered
A stored memory can be retired from recall or erased outright. Data storage and privacy covers the difference and when each one is the right call.
For a lesson that turned out to be wrong, the usual move is neither: write the correction, and the old lesson is superseded and stops being served, with its history intact. Agent memory poisoning and integrity covers how a contested correction is settled.
For larger teams
Access by role, identity from the directory you already run, and an auditor's view of who could see what are enterprise controls. AI agent governance and audit trail covers them.