Multi-agent memory: one person, many agents
Everyone is building an agent. The person using five of them is the one nobody is building for.
Configure · · 4 min read
In short
- Multi-agent memory can mean agents inside one system sharing state, or one person's memory shared across separate agents from different companies.
- The first is an engineering problem inside your own product. The second needs infrastructure between products.
- People will use many agents by default, because specialized agents will usually be best at their own product's tasks.
- Shared memory across companies needs identity, permissions and a neutral operator, with the user in control.
Multi-agent memory means memory that more than one AI agent can read and write. The phrase covers two different things. One is a team of agents inside a single system sharing notes while they work on a task. The other is one person's memory being available to separate agents built by different companies. The first is common. The second is mostly missing.
We build Configure, infrastructure to connect agents, apps and chats, and the second meaning is our whole job. This article explains both so you can tell which one you need.
What are the two kinds of multi-agent memory?
| Inside one system | Across separate agents | |
|---|---|---|
| Who the agents are | A planner, a researcher and a writer that you built | Assistants and apps from different companies |
| Whose memory | The task's working state | A person's preferences, history and connected data |
| Who controls it | You, the developer | The user, for each agent |
| Main problem | Keeping agents consistent and not overwriting each other | Identity, consent and trust between companies |
| Usual solution | A shared store or message log in your framework | A neutral layer with permissions |
If your agents all run inside your own product, you have the first kind, and your agent framework or a memory library is the right tool. You do not need anything from us for that. The rest of this article is about the second kind.
Why will one person end up with many agents?
Because everyone is building the same thing from a different starting point. More companies now offer a personal agent attached to an ecosystem they already own: one in a social network, one over messaging, one connected to your mail and documents. At the same time, individual services are building specialized agents for their own products. A food delivery app has one for ordering, a shopping site has its own assistant, and a travel site has one for booking.
Those specialized agents will likely always take priority inside their own platforms. They have capabilities there that an outside agent does not, and they will usually be best at their own product's tasks. So we expect people to use many agents by default, even if they have a preferred one. A travel plan may begin in your favorite assistant and end inside an airline's product.
There is also a simple matter of numbers. If we stay on a path where there are more agents than people, nobody will be able to build compatibility with each of them by hand.
What breaks when those agents share nothing?
Every handoff becomes a restart. The dates and budget you settled with one agent have to be typed again for the next. Each service builds its own partial picture of you, and none of them match. The user does the integration work in their head, carrying facts from one conversation to another.
The impact of AI is much larger if intelligent applications can operate as an ecosystem. Picture someone who planned a trip with an assistant. They let a travel site use the dates and budget already agreed, then share the itinerary with the airline, the hotel and a restaurant booking app as they book. Each service receives only what the user approves, and all of them work from the same plan as it develops. That is what shared memory across agents is for.
What does shared memory across companies need?
Agents do not need the bespoke wiring that older software integrations did, because they work in natural language. They still need infrastructure for three things.
- Identifying the same person across services. Two products know you by two accounts. Something has to connect them without exposing either product's users to the other.
- Maintaining the connections. Memory changes with every conversation. A one-time copy is stale immediately, so the link has to stay live.
- Enforcing what each agent can access. The user approves each agent, can narrow what it reads, and can revoke it. Every read is checked first.
Configure handles those three. Everything a user connects is linked to a single profile they own. Each agent writes only to its own namespace, so the source of every memory is known. Before another agent reads anything, the request is checked against the user's permissions. We describe the mechanics in Cross-agent memory: how agents share what they know.
Why can't one of the big assistants just be the hub?
Each of them would like to be. The trouble is that the other parties have to agree, and they will not agree to a direct competitor holding their users' context. Connecting information across counterparties has historically been done by a neutral party for exactly this reason. Payment networks and bank-connection providers exist because banks and merchants were never going to coordinate fairly on their own.
For agents, neutral means having no frontier model and no consumer platform to favor. An independent operator is the arrangement every side can accept, and it is the only one in which a specialized agent and a general assistant can both take part on equal terms.
How do you know which kind you need?
- Your agents are all yours and work on one task together: use your framework's shared state. That is multi-agent memory inside one system.
- Your one agent forgets users between sessions: you need a memory layer in your own product.
- Your users arrive already known to other assistants and you start from nothing: you need memory across agents. The quickstart shows how to add it, and the MCP overview covers assistants that connect through tools.
Many products end up with two of the three. They are separate problems, and it saves time to name which one you are solving before you pick a tool.
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