How it fits

Agents come and go. The context stays.

Information lives where it always has. Agents and models change every few months. Governed context sits between them: the one place that decides what any agent may know, keeps it current, and records every read.
The stack

Three layers. Only one of them should hold the truth.

Layer 3
Agents, models and runtimesSwappable. A run can end when its job is done.
1Claude3CursorPromptOwl agents2Scheduled job on a small office machineCloud runner
Read, scoped by rolePropose updates, an owner approves
Layer 2
Governed context layerContextNest. Stays put.

The company's knowledge, in plain files it owns, that every agent reads.

Approved7Current4Scoped to each jobVersioned5Every read recorded
Distill and approve what agents may use
Layer 1
Where information livesStays where it is. Agents don't read it directly.
6Google DriveSharePointEmailCRMMeeting transcripts

Where each problem shows up, and where it's fixed

  1. 1
    Two agents give different answersEach agent keeps its own copy. Fixed when every agent reads one approved layer.
  2. 2
    Brilliant on a laptop, dumb in productionContext was stuck on one machine. Fixed when context lives outside the runtime.
  3. 3
    Switching models means starting overContext sat inside one vendor's tool. Fixed by plain files every model reads.
  4. 4
    An agent reads what it shouldn'tHappens when agents reach sources directly. Fixed by giving each agent only the knowledge its job needs.
  5. 5
    Nobody can say what the AI was working fromFixed by recording every read (person or agent, document, time) and keeping every approved version.
  6. 6
    Search gets worse as the drive growsAgents searching raw storage. Fixed by distilling an approved set.
  7. 7
    Answers from last year's policyOld versions sit next to new ones. Fixed when agents read the current approved version.
Watch it happen

One person's AI, then everyone's, then governed.

Step through it: one agent works fine, four agents disagree, then one governed layer gives every agent the approved version.
Refund policyRefund policy 2023Pricing sheetPayrollYour Claude✓ Refunds: 30 days

One person, one agent, their own files. It works, because it's yours.

Different layer, different job

Memory sits inside one agent. Context sits under all of them.

Memory tools are useful and increasingly free. They remember things for one app or one user. Governed context is shared company knowledge with owners, approvals and a record of every read. You'll often run both.

No layer

Every agent wires into every source with its own permissions and its own idea of what's current. Three agents and four sources means twelve connections, and it grows with every agent you add.

Memory tools

Each agent remembers what it learned, written by the agent itself. The memories aren't shared, nobody approves them, and they stay inside that app.

Governed context layer

Sources feed one layer. Every agent reads it, scoped to its job. Add an agent and it starts with everything it's allowed to know.

AI memory vs governed context, in detail →

Where to run agents

Run them wherever they're cheapest and most capable.

Run the CRM and project boards anywhere ordinary web software runs. Run long agent jobs wherever they work best: a laptop, a small always-on machine in the office, or a cloud runner. Because the context lives in the layer, moving a job doesn't lose anything.

That matters because an agent on your own machine is unusually capable: it has your logins, your files and no time limit. Boxed into a web app with narrow credentials and a metered bill, the same agent gets much less done. Keep the knowledge portable and you can pick the runtime job by job.

You already budget for identity, security and data layers.

This is the layer that decides what your AI may treat as true. Few teams have a name for it yet, even when they're living with the problem, so start from the symptom.

Find yours →
In the product

Owners approve. Agents read the approved version. Every read is on the record.

Approving a change to a deal desk policy in ContextNest, with Approve and Reject buttons
Approve or reject with the change in view. Agents keep reading the approved version until then.
ContextNest activity log for one document, listing reads over MCP and the web, an edit made through Claude Code, and a review request
The record for one document: who read it, over which surface, and every change, including the one an agent proposed.

Screenshots from ContextNest Community 1.28, using a fictional demo company.

“The great thing about working with the PromptOwl team is that they're willing to dance. We don't know exactly what we want because AI in education is so new — but they adapt with us. You don't find that kind of partnership very often. It's rare.”

Dennis DiNoia, Founder of Mr. D Math
Dennis DiNoia, Founder, Mr. D Math10,000+ students, 15+ years of teaching expertise now built into a governed AI tutor.Read the full story →
Why teams pick us

Why teams pick ContextNest

Governance that feels like doing your work. You write a doc, chat with Hootie and run your agents, and the approvals, versions and read records happen along the way. Hyperscaler governance is IAM policies, pipelines and consoles you operate.

Why we built ContextNest →
Runs anywhere. Same nest.
Start with plain files on a laptop with the open-source CLI, self-host the server in Docker, or let us host it. Push the same nest from one to the next, with no migration.
No infrastructure to stand up
No graph database, no vector cluster, no data platform, no ontology pipeline. It starts as plain files, and hosted is free to start.
A product people open every day
A Notion-style wiki editor, approvals, version history and compare, a read log, and Hootie chat built into the nest. PromptOwl agents run on top.
Agents that keep it whole
Capture, curator and retriever agents ship in the ContextNest plugin for Claude Code: they keep what's worth keeping, make a correction land everywhere, and answer with citations.
Governed and enterprise-ready
Turn on review and a named owner approves before AI may use it, every version is kept, and every read is recorded. OIDC single sign-on with Entra ID, Okta or Google Workspace comes with Enterprise.
Open where it counts
The CLI, engine and MCP server are open source (AGPL-3.0), with an open spec and free starter vaults. It all runs on a laptop.

Put the knowledge outside the agent.

Connect the AI tools you already use to one governed layer, in minutes.

Get PromptOwl freeGet a demo