Xerum compared with an agent memory layer
Memory remembers what your agent said. Context records what your team knows.
These are not competing implementations of the same idea. A memory layer gives an assistant continuity across its own sessions. Xerum is the team's record — pulled from where the work happened, versioned, and curated by people. If you want both, you can have both.
What a memory layer is genuinely good at
Continuity, cheaply
An assistant that remembers your preferences, your codebase quirks and what you asked yesterday is meaningfully better than one that does not — and a memory layer delivers that with almost no setup.
It requires no source of truth
There is nothing to connect and nobody to approve anything. It learns from the conversation it is already in, which is why it works on day one and why teams reach for it first.
Genuinely hard engineering
Deciding what is worth keeping, promoting facts between scopes and tracking when a fact stopped being true are real problems, and the good ones solve them well.
The same question, in each
Where does a fact come from?
A model inferred it from a conversation with a user, and stored it.
An author wrote it at a source. A connector recorded it as a resource carrying its origin — source, author, URL, time.
Who can change it?
The agent, silently, whenever it decides something has changed.
Nobody, directly. A member or an agent opens a proposal; an owner approves it or rejects it.
What happened to the old version?
Usually overwritten. Some layers track validity intervals; most mutate the fact in place.
Kept. Nothing is edited in place — every change appends and makes a new revision you can read at.
Can I check where an answer came from?
Back to the agent's own earlier conversation, at best.
Back to the origin URL, via the revision and resource. The chain ends at a human's words.
Who else can read it?
Scoped per agent or per user. Sharing across a team is usually not the model.
The whole team, filtered per person inside the query, and every agent any of them runs.
Where does it physically live?
A multitenant API, typically, with your data alongside everyone else's.
Your own container, your own database, your own storage prefix.
Keep your memory layer
Nothing here asks you to remove one. Session continuity is a real problem and Xerum does not solve it — an agent that forgets your working style between prompts is still annoying whatever it can cite. What changes is where the shared, durable, team-level facts live: not in a per-agent store that a model wrote and nobody reviewed, but in a record with origins, revisions and an owner who approved it.
Give your agents something worth remembering
Built today, not on a roadmap. Every property above is already how the store works.