Customer AI assistant
Assistants that already know you
Four assistants built on the same memory layer. Each one keeps what it learns — your preferences, your team's decisions, your codebase — so it stops asking twice.
The problem
A blank slate every morning
Most assistants are impressive for one exchange and exhausting by the tenth. You explain your format preference again. You paste the same background again. You correct the same mistake you corrected last week. That isn't a model quality problem. The model is fine. It just isn't given anything it learned yesterday, because nothing kept it.
The four assistants
Pick the one that matches the work
Work assistant
Shared team memory — decisions, context, and documents that outlive the thread they were discussed in. Link See the work assistant →
Personal assistant
Learns one person over time. Tone, format, preferences, and past corrections persist automatically. Link See the personal assistant →
Content generator
Drafts in your voice, because it remembers every edit you've made to its previous drafts. Link See the content generator →
Code generator
Suggestions grounded in how your repository actually works, retrieved at method and class level. Link See the code generator →
Shared foundation
One memory layer underneath
All four assistants read and write to the same store, so context earned in one surfaces in another. A decision recorded by the work assistant is available to the content generator writing the announcement about it. Scoping keeps that from becoming a leak. Personal memory stays personal; team memory is shared within the team; account memory spans the organisation. You set the boundary when you write.
Deployment
Self-hosted, so it can hold real context
An assistant is only useful if it can remember things you'd otherwise be careful about sending anywhere. Zenmem runs entirely on your infrastructure — vectors, documents, and memory never leave your network, and there's no per-call metering watching how much you use it.
FAQ
Can I use more than one assistant?
Yes, and it's the intended setup. They share the memory layer, so context compounds rather than fragmenting.
Can I customise how an assistant behaves?
Each assistant exposes behaviour, tone, and retrieval settings. If you need more control than that, the agent builder lets you compose one from scratch.
Does the assistant improve over time?
Retrieval does. Zenmem tracks which memories the model actually used and tunes what surfaces on similar queries, so the assistant gets more relevant with traffic rather than noisier.