Memories
Think of a memory as a shared notebook for your agents — except this notebook files itself. Write down "Sarah at Acme prefers short emails," and it doesn't just save the sentence: it notices Sarah, notices Acme, and remembers that she works there. Next time any agent drafts an email to Sarah, it already knows.
Without a memory, every conversation starts from zero. With one, your agents build up context the way a colleague does — a little at a time, and they never forget it.
The difference, in one conversation
Same agent. Same question. The only thing that changed is whether it had somewhere to remember.
You mentioned: "Sarah Chen runs ops at Acme. She prefers short emails and hates attachments — send links instead."
Happy to help. Who is Sarah, and which renewal are you referring to? Would you like me to attach the proposal document?
The conversation ended, and so did everything it knew. You re-explain from scratch — every time.
Drafted a short note to Sarah on the 12-month renewal she owns — no attachment, the rate card is linked inline. Want me to send it?
It checked the memory before writing a word, and applied a preference you only mentioned once.
Notice what it did not have to ask. The memory carried who Sarah is, what the renewal is, and how she likes to be written to — across three weeks and two unrelated conversations.
What you can do
| You can… | What that means for you |
|---|---|
| Keep facts across conversations | Your agent remembers last month's decision in today's chat |
| Let agents write to it themselves | It records what it learns while working, without you dictating |
| Share one memory across many agents | Your sales agent and your support agent see the same customer history |
| Use the same memory outside Nirvai | Claude Code, Cursor or Codex read and write the very same notes |
| Keep the original documents | The contract behind the note stays attached, ready to re-read |
| Correct or forget anything | Facts change — you can edit or delete them, and the links update |
How a memory fills up
In a chat, a task, or a document your agent just read.
One small, self-contained note — a single fact, in plain words.
The note names who and what it involves, and how they relate.
Before acting, any connected agent looks it up — and finds the whole thread.
The pieces have names you'll see throughout the app: an observation is one fact, the people and things it mentions are entities, and how they connect are relationships.
One fact per observation. "Acme renewed for 12 months" and "Sarah owns the renewal" are two notes, not one — so either can be corrected later without touching the other.
What a memory looks like
Each memory is its own store, with its own subject. Most people keep a few: one for clients, one for how their team works, one personal.
My clients and where each stands, key contacts, and how I like email handled. Consult before drafting or triaging.
The description is the most important field — it's what an agent reads to decide whether this memory is worth consulting.
"Shared" means everyone in your organization can use it. Without it, the memory is yours alone.
The counts tell you how much it knows: notes, the people and things in them, and the connections between those.
One memory, every agent
A memory isn't tied to a single agent. It's a store you connect things to — so the same facts serve every agent you use.
- Agents you build in Nirvai. Connect a memory while building an agent, and choose how much it may do. See creating your agent.
- Agents you already use elsewhere. Claude Code, Cursor, Codex and other assistants reach the same memories through your Nirvai connection — so a fact your Nirvai agent recorded this morning is there when you open your terminal this afternoon.
That's the point: the memory is the shared part. Switch assistants, add a second agent, work from a different tool — the accumulated knowledge stays put.
| Access level | What the agent may do |
|---|---|
| Read only | Look things up. Never changes what's stored. |
| Read & write | Look things up, and record new facts and files. |
| Full access | All of the above, plus correct and delete existing facts. |
New connections start at Read & write. Give Full access only to agents you trust to fix their own mistakes.
A knowledge base is reference material you upload for an agent to read — a manual, a price list. A memory is something your agents write to as they work, and it's shared between them. Use a knowledge base for what you already know; use a memory for what your agents learn.
What's next
Create a memory
Describe it in plain words and let AI draft it, or start blank.
Explore what it knows
Read the notes, browse the files, and walk the connections.
Connect an outside assistant
Reach the same memories from Claude Code, Cursor or Codex.
Technical reference
The data model, recall behavior and access tiers.