Xerum compared with a platform's own memory
Your team runs more than one vendor’s agent
Anthropic, Google and OpenAI all shipped memory in 2026, which settled the question of whether agents need persistent context. What they shipped is scoped to one assistant and locked to one vendor — so the second agent your team adopts starts from nothing again.
What platform-native memory is genuinely good at
Zero setup, and it is already there
It arrives switched on inside a product your team already pays for. Nothing to connect, nothing to provision, no procurement conversation.
Deeply integrated with one assistant
Because the vendor controls both halves, the memory is used well — at the right moment, in the right context window, without you orchestrating any of it.
It validated the whole category
Three platforms shipping the same capability in one year is the strongest possible signal that agents need durable context. We are arguing about the shape, not the need.
The same question, in each
Which agents can read it?
That vendor's, and only that vendor's.
Any agent that speaks MCP — Claude, Cursor, and whatever the team standardises on next.
Who does it belong to?
A person and their assistant. Team-level sharing is limited or absent.
The team. One shared context, filtered per person at query time.
Who wrote what is in it?
The model, from your conversations with it.
Authors at sources, via connectors — or people, via proposals an owner approved.
Can I audit it?
You can usually view and delete entries. There is no history of how they changed.
Every change carries author, time and message, and no revision is ever removed.
What happens if you switch vendors?
The memory stays behind. It is part of the product you are leaving.
Nothing happens. The context is yours, in your instance, and the new agent connects over the same protocol.
Where does it live?
The vendor's infrastructure, under the vendor's terms.
Your own container, database and storage prefix.
Keep using it
Per-assistant memory makes that assistant better at being itself, and turning it off would not benefit anybody. The gap it leaves is the one that matters at team scale: your engineers use one agent, your support org another, and your analysts a third, and none of them can read what the others learned. The shared, durable, cited layer has to sit outside all three — which is the only place it can be vendor-neutral.
One context, every agent you will ever run
Built today, not on a roadmap. Every property above is already how the store works.