Skip to content

Context compounds: an AI context layer for teams

Every AI tool your team uses starts from zero. A shared context layer gives people and agents one place to work from, governed by your company.

Jacob Cansler · Maker of Gryz

Published Updated 6 min read

For your agent
For your agent

Short answer: An AI context layer gives your people and agents one place to work from. Gryz is that shared context: memory, documents, tasks, skills, notes and search, for Claude, Claude Code, ChatGPT, Cursor, Codex and any MCP app, with a company workspace that decides who reaches what.

Models are becoming commodities. Context isn't.

Your team probably uses three or four AI tools already. Someone plans in ChatGPT, someone writes in Claude, engineers live in Cursor or Claude Code. Each one is smart. Each one also starts every chat knowing nothing about your company.

So people paste the same brand guide, the same spec, the same "here's what we decided last week" into every new chat. The model does fine work. Then the chat ends, and the context goes with it.

That is the part worth fixing.

Why context compounds and models don't

A better model shows up every few weeks. You can switch to it in an afternoon. What you can't download is everything your company knows:

  • who the people and agents are, and what each one owns
  • what you are working on, and why
  • what was decided, and what was ruled out
  • who owns the next action
  • who is allowed to see or change what

A model gives you the same answer for everyone. Your context is the part that's only yours, and it gets more valuable every week you keep it. Lose it at the end of each chat and you start from zero again.

We wrote about this for one person in Agent memory that works across every model. For a business the stakes are bigger, because the context belongs to a whole team.

What a company context layer holds

Gryz is the shared context your people and agents work from. In a company workspace, it holds:

  • People and agents. Members with roles, and company agents with their own names, keys and permissions.
  • Projects. Each one can be open to everyone or restricted, and can have a short North Star your people write: who it's for, what matters most, what not to do.
  • Memory. Facts and decisions saved by people and agents, with who saved them. A project can hold agents' new memories until a person approves them.
  • Documents. Specs, guides, briefs and plans, with version history and suggested edits that a person accepts.
  • Tasks. Owned work with a status, an assignee and, if you want, a named reviewer. See Company tasks.
  • Skills. Shared know-how, like a review checklist or a writing style, that owners publish for agents to read. See Company skills.
  • Search. One search across the company's documents, memories and tasks, by meaning and exact words. People and agents only get results they could already open.

None of this is a pile of files. It's connected: a task links to documents in its project, a memory sits in the project it's about, and access follows the project.

Any AI tool, one connection

Gryz speaks MCP, the open protocol most AI tools now use for outside context. That means the same context reaches Claude, ChatGPT, Cursor, Claude Code, Codex and any other MCP client.

There are two ways in:

  • Connect an app once. Connect Claude, ChatGPT or Cursor to your own Gryz, and it reaches your Private space plus each company you add to it. Every call works in exactly one workspace, and it starts read only in a company.
  • A company agent with its own key. An agent that lives in one company and uses only the projects you give it. Good for a bot that should do one job and nothing else.

Agents come and go. Gryz remembers. When your team switches tools, the context stays put. More on this for developers in Shared context for AI coding agents.

Governed by the company, not by a prompt

A shared context layer is only useful if the company decides who reaches it. In Gryz that control lives in settings, not in an agent's instructions:

  • Connected apps are off until an owner turns them on. Then the owner picks a policy: Team (Claude, ChatGPT and Cursor allowed, other apps need approval) or Enterprise (every app needs approval, with stricter protections on). See Team and Enterprise.
  • Projects decide access. A restricted project stays inside its people. Mark a project No connected apps and no app a member connects can see it.
  • Agents propose, people decide. An agent never closes a task: it goes to Needs review. Agents can suggest edits, profile changes and North Star changes, and a person approves them.
  • Activity records what happened, like which agent used which tool in which project, without recording the text.
  • No public links. While an app can see a company, it can't publish anything publicly.

Private and Company, kept apart

Everyone in a company has two spaces:

  • Private is yours: your documents, notes, memory, tasks and skills. Owners and admins can't see it.
  • Company is the shared space, governed by roles and projects.

Nothing moves between them by itself. When you're ready, Bring your work in shows you a plain summary of what will move and which links will end before anything happens. The full list of who sees what is in What is private in a company workspace.

A day with a shared context layer

Here's how it can look for a small product team. (A made-up example, not a customer story.)

  • On Monday the product lead decides, in ChatGPT, to push a feature to next month. She asks it to save that decision to the Launch project's memory.
  • On Tuesday an engineer opens Cursor. His agent recalls the project memory, sees the decision, and doesn't start on the feature.
  • A company agent working the Launch project picks up a task to update the release notes. It edits the document, and the change is saved as a new version with the agent's name on it. It marks the task done, which sends it to Needs review.
  • The product lead approves it from her Review inbox.
  • Marketing asks Claude to draft the announcement. Claude searches the company's documents, finds the brand voice skill and last quarter's launch post, and writes on-brand.

Nobody pasted anything. For the team side of this, see An AI knowledge base your marketing team owns.

Questions people ask

What Gryz is not

Honest limits, so you can decide if it fits:

  • Not an agent. Gryz owns context, not execution. Your agents run wherever they run. Gryz gives them memory, access and a shared record.
  • Not in control after an app reads. When Claude or ChatGPT reads company text, it goes to that provider under your own agreement with them. Gryz decides what an app can reach and records it, but can't control what the app does with the text afterward. AI apps and your company covers this plainly.
  • Search sends text to Cloudflare. To make documents searchable, Gryz sends short passages to Cloudflare, our AI provider. Company search is off until an owner or admin turns it on. See How search works.
  • Hosted in the United States. Gryz stores your data in the US.
  • Rolling out. Company workspaces and connected apps are rolling out, so some of this may not be on for your company yet.

Get started

Your team will keep changing tools. Make sure the context stays.

More on these topics: Shared context for AI agents and Search your documents from any assistant.

Was this page helpful?
  • Gryz vs Mem0: which one should you use?

    Looking for a Mem0 alternative? Mem0 is a memory API for apps you build. Gryz is shared context for your own agents: memory, docs, tasks, skills and search.

    Comparisons5 min

  • Chat with your documents from any AI tool

    Import your Word files and notes, then ask ChatGPT or Claude about them. Gryz makes your library searchable from any AI tool, privately, with no setup.

    Document search4 min

  • Shared context for AI coding agents, over MCP

    Claude Code in the morning, Cursor after lunch, Codex at night. One MCP connection gives them all the same memory, docs, tasks, skills and search.

    Developers6 min