Blog

Cross-agent memory: how agents share what they know

What you told one agent should be able to help the next one, if you say so. That is cross-agent memory.

Configure · · 5 min read

In short

  • Cross-agent memory means an agent can read what other agents have learned about the same person, with that person's permission.
  • Cross-agent context is the wider term: memory plus connected app data and anything else an agent needs to pick up where another left off.
  • The hard parts are knowing it is the same person, tracking where each memory came from, and enforcing what each agent may read.
  • It works when a neutral layer sits between the agents and the user controls every connection.

Cross-agent memory is the ability of one AI agent to use what another agent has already learned about the same person, with that person's permission. A memory saved in one product can inform a different product from a different company. The user decides which agents take part, and each agent only sees what it has been approved to see.

We build Configure, infrastructure to connect agents, apps and chats, so this is the problem we work on every day. Here is how we think about it.

What is the difference between cross-agent memory and cross-agent context?

The two terms get used for the same thing, and they overlap, so it helps to separate them.

Cross-agent memoryCross-agent context
What it holdsFacts and preferences agents saved about a personMemory, plus connected app data and the state of the task in progress
Example"Prefers TypeScript. Vegetarian. Training for a marathon."The trip plan agreed yesterday, this week's calendar, the thread with the client
LifespanLong. It stays true for monthsMixed. Some of it expires when the task is done
Who writes itAgents, as they learnAgents, apps, and the user's own imports

Memory is the durable part. Context is everything an agent needs in order to continue an experience that began somewhere else. In practice you want both to travel together, which is why we treat memory as a byproduct of connecting agents and not as the product itself. We wrote more about the wider category in What is context as a service?

Why does memory need to cross agents at all?

Because people will use many agents by default. Companies are attaching personal agents to the ecosystems they already own, and individual services are building specialized agents for the work inside their own products. Those specialized agents will usually be the best at their own tasks. So even someone with a favorite assistant ends up talking to several others.

Each one starts from nothing. You explain your diet to a recipe agent on Monday and to a restaurant agent on Friday. You describe your codebase to one coding agent and then again to a second. The information exists. It just lives in a place the next agent cannot reach.

Technical users who are ahead of the curve already fix this for themselves. They wire up a personal agent setup, sync memory between tools, and spend hours maintaining it. Most people will not do that, and should not have to.

What makes cross-agent memory hard?

Agents work in natural language, so they do not need a shared data model to understand each other's notes. That removes the old integration problem. Three harder problems remain, and all three sit below the agents.

Knowing it is the same person

Two products know you by two different accounts. Before any memory can cross, something has to establish that both are you, and it has to do that without either product handing over its user list. In Configure, a person verifies once and that verification is what links their profile across the agents they approve.

Knowing where each memory came from

A fact with no source is hard to trust and impossible to manage. Configure records which agent wrote each memory and keeps every agent's contributions in its own namespace. An agent can write only to its own space. That is what lets a user look at their profile and see what each agent has added, and remove any of it.

Enforcing who may read what

Sharing has to be the user's decision, made per agent. Every read is checked against the user's permissions before anything is returned. If the user has not approved an agent, that agent gets nothing. If they revoke it later, its next read is refused. The details are on our trust page.

How do agents share memory today?

How it worksWhere it falls short
Copy and pasteThe user pastes a summary from one assistant into anotherManual, one time, and out of date by the next conversation
Each assistant's own memoryAn assistant remembers you inside its own productStops at the edge of that product
A personal setupA technical user syncs files or tools between agentsHours to build and maintain. Not for most people
A neutral memory layerAgents read and write a profile the user owns, with permission checksEach product has to integrate once

The first three all work for somebody. Only the last one works for an ordinary user across products they did not set up themselves.

Why does the layer in the middle have to be neutral?

Connecting information across counterparties has historically been done by a party that is not one of them. An AI lab has a model to favor. A platform has an ecosystem to favor. Neither can be expected to coordinate fairly with its direct competitors, and nobody should want them to hold everyone else's users' memories.

An independent operator with no frontier model and no consumer platform of its own is the arrangement all sides can accept. It also has to be more than a pipe. A tool that only pulls memory out of one assistant is exfiltration with a friendlier name. A real layer serves and syncs memory in both directions, with explicit user consent at every connection.

How do you add cross-agent memory to a product?

With Configure there are two ways in, depending on what you are building.

Either way, the first sign it is working is simple. A user opens your product for the first time, and the agent already knows something true about them that they told a different agent last week.

Add a connect button to your product

The quickstart takes one page, end to end: a chip in your app, Configure's hosted page, and your agent reading the user's context on the first turn.

Read the quickstartNext: AI onboarding without the questionnaire