We didn't just change labels. We changed what Norfolk is.

---

For a while, Norfolk was easy to explain. A notes app. Capture stuff, organize it, find it later. Simple.

That explanation was useful. It was also incomplete.

Because while we were building it, Norfolk quietly became something else: a place where agents can read, recall, and write memory — with provenance, scope, and structure.

Not "chat with your notes." Not RAG over a folder.

A memory surface.

The old question was wrong

When Norfolk was just "a notes app," the obvious question was: Why not Obsidian? Why not Notion?

Fair question. Wrong framing.

Norfolk is not trying to win by being a better editor. Its advantage is what happens after capture.

You drop in raw material — screenshots, PDFs, rough notes, links, half-formed thoughts. That part feels like any notes app. But the material doesn't just sit there. It becomes recallable context. An agent can find it, cite it, build on it.

Notes become working memory.

That's the difference. And we should have said it earlier.

What actually changed

Two user paths are now explicit.

Personal Norfolk — Your agent can read from and write to your Norfolk space directly. Your notes stop being isolated from your AI workflow. A coding agent, research agent, or work assistant can use your Norfolk as context.

But cross-recall is off. Norfolk and Nexus don't merge into one search surface on personal plans. The agent works with Norfolk, but each store stays separate. You pick which one to talk to.

That's intentional. Direct connection, not full shared infrastructure.

Nunchi Team ($29/month) — This is where it gets interesting.

Norfolk and Nexus become one recall surface. What a human captured in Norfolk and what an agent remembered in Nexus are searchable together. Shared quota. Shared recall. 120,000 MU across both.

Why does this matter? Because teams don't fail from lack of a better model. They fail because every person and every agent keeps rebuilding the same context from scratch. The PM wrote the decision in Norfolk. The developer's coding agent has no idea it exists. So the developer asks on Slack. Or worse, the agent guesses.

Team memory fixes that. One recall call, both stores, provenance preserved.

The architecture behind it

Here's the mental model.

Nexus is the agent memory layer. Sessions, decisions, work history, execution context. What the agent remembers.

Norfolk is the human knowledge layer. Notes, documents, research, uploads, agent outputs worth keeping. What the team knows.

AMCP is the protocol that lets any agent work with either store through the same interface — recall, remember, sessions, export.

Claude Code / Cowork / Your Agent
              ↓
       AMCP-MCP Server
              ↓
        Policy Layer
         ├── Nexus   (remembering)
         └── Norfolk (knowing)

The agent doesn't need to know the storage details. It calls recall with a target. The server handles routing, scope, and access control. On team plans, target=auto searches both stores and returns results tagged by source.

Every result carries provenance: who created it (human or agent), which store it came from, which session it belongs to. No ambiguity about what's a human decision and what's an agent's working note.

Why not just a wiki? Why not just files?

There's a popular idea right now: give your AI a folder of markdown files and let it maintain a wiki. Read sources, update pages, cross-reference.

The insight is right. Knowledge should compound, not restart every session. RAG gives you your principal back each time. A wiki gives you interest.

But file-based wikis hit a ceiling fast.

When you have hundreds of files, the agent needs an index file to know which files to read. That index needs maintenance. When the index drifts, the whole wiki becomes unreliable. You built a system to reduce maintenance cost, and you got... another maintenance problem.

And that's before you ask harder questions. Can two agents share the same knowledge? Can a team member's notes become recallable by someone else's agent? Can you tell whether a piece of knowledge was written by a human or generated by an agent?

Files can't answer those questions. A protocol can.

That's why memory needs to be infrastructure, not a folder.

Our take on memory

A lot of AI products describe memory in vague ways. Chat history. Folders and tags. Hidden retrieval glued inside a harness. A feature of the model itself.

We think memory needs sharper properties.

Memory must outlive the session. If it disappears when the tab closes or the runtime restarts, it's temporary context, not memory.

Memory must carry provenance. An agent should know not just what was remembered, but where it came from, who created it, and what scope it belongs to.

Memory must be usable by more than one actor. A human's note shouldn't be trapped in a human-only UI. An agent's memory shouldn't be locked inside one vendor's runtime.

Memory must be portable. Models improve. Harnesses change. Vendors change. Memory should survive all of that. That's why AMCP exists as an open protocol — so memory isn't coupled to any single tool.

Norfolk is still a notes app

This doesn't mean Norfolk stops being easy to use.

The opposite, actually. The capture experience matters more now, because the human side of memory has to stay natural.

People don't start with schemas. They start with a screenshot. A rough thought. A link saved too early to classify.

Drop it in. Don't force structure. Don't make the human become a database admin.

Norfolk stays simple at the front. But the backend meaning of that captured material is now bigger. What starts as a note can become recallable context. What starts as personal capture can later participate in agent workflows.

We're not moving away from notes. We're giving notes a larger job.

Where this lands

From the user side, the story is now cleaner:

The industry keeps asking which model will win. We think that's often the wrong question.

Models will improve. What matters longer is whether memory survives those changes.

That's what we're building.

---

Norfolk is part of Nunchi AI's memory infrastructure. Try it at nunchiai.com.