What Is a Context Layer? A Plain-Language Guide

In the previous piece we argued that your AI tools underperform because they have no memory of your business. The fix is a context layer. This piece explains what that is, in plain language, and in enough detail that you could sketch it on a whiteboard afterward.

One sentence first

A context layer is one memory of your business that every AI tool you use can read from and save to.

Not a chatbot. Not another app your team has to live in. A layer that sits underneath the tools, the way a database sits underneath your website. Your AI tools do the thinking. The layer remembers. Your team mostly notices that every AI they touch suddenly knows the company.

The three moving parts

1. AI tools save to it. The AI tools your team already uses, ChatGPT, Claude, Claude Code, Cursor, connect to the layer once. From then on, as they work, they save what matters: the decision made on a call and why, the commitment a client is waiting on, a new contact, the note that explains how an account really works. Where the work involves another app, a calendar, a CRM, a transcript tool, the AI tool reaches that app through its own connectors, does the work, and saves what is worth keeping. You can also drop files in directly: documents, a whole folder, a zip. Nothing about how your team works has to change much, which is the design constraint that makes adoption real. Systems that ask humans to stop and type starve; the layer takes dictation from tools that were in the room anyway.

2. The layer keeps itself in order. This is the part that separates a context layer from a pile of notes, and it deserves its own paragraph. Every memory has a type: a note, a decision (which must carry its reason), a task (open or done, with an owner and a due date), a person, a customer, a meeting, a file, or part of your identity, such as your values or your voice. Memories link to each other, so a customer connects to their people, their meetings and the decisions made about them. The layer writes Summaries: a plain account of each day, rolled into weeks and months, so the working picture stays readable instead of growing into a swamp. Identical duplicates are merged with both histories kept; similar ones are put to you as a question. Every change records who made it, with which tool, and what it was before. The next piece unpacks the structure itself; it is simpler than you would expect.

3. AI tools read from it. Any AI tool that speaks MCP, the open standard for this, can ask the layer: who is this person, what is open with this account, what did we decide about pricing, what changed this week. At the start of a session it gets a brief: what the memory is for, what happened recently, and what changed since it was last there. Searches work by meaning, not just keywords. And because there is one memory, not eight, asking the same question from two different tools gets you the same answer.

What a brief looks like

Here is an illustration of the kind of brief an AI tool receives when a session starts. The names are invented; the shape is the point.

Engine: Client work
Purpose: Everything about our client relationships: who they are,
what we promised, what we decided. Not internal HR or finance.

Recently
- Meridian Freight quarterly review (Tue): agreed to move to annual
  pricing; Sarah asked for a revised proposal by Friday.
- Decision: annual-only pricing for new customers (rationale saved).
- Task: send Meridian revised proposal, owner Deniz, due Fri.

Since you were last here
- 3 new memories from ChatGPT · Deniz
- Task "Meridian onboarding checklist" marked done

The tool starts the conversation already knowing this, without you pasting anything.

A day in the life, concretely

Tuesday, 8:30 a.m. You open Claude and ask it to prepare you for today’s calls. Claude reads your calendar through its own calendar connector, then looks up each attendee and account in the memory: who they are, where things stand, what is open. Nobody assembled that by hand. (The pre-meeting briefs piece goes deeper.)

10:00. The client call happens. Afterwards you hand the transcript to your AI tool and say “save what matters.” It saves the meeting, the decision made on the call with its reason, and the two commitments, yours and theirs, as tasks with owners and due dates.

10:45. Your colleague asks ChatGPT for the follow-up email. ChatGPT already knows both commitments and the decision, because it reads the same memory Claude just saved to. The draft is right the first time.

6:00 p.m. The Summaries tab has the day written up: what moved, what was decided, what is open going into tomorrow.

No one maintained a knowledge base at any point. The tools did the work, saved as they went, and the layer kept it in order. Multiply Tuesday by a year and you have the compounding asset: every day of operation makes every future task start from a richer picture.

What a context layer is not

Precision matters here, because several things wear the word “memory”:

• It is not the memory feature inside ChatGPT or Claude. Those are personal and vendor-locked: they remember your preferences, inside one product. A context layer is shared and neutral: it remembers the business, for every tool.
• It is not enterprise search. Search finds documents when you ask. A context layer holds a structured picture and hands your AI tools the slice they need. The difference shows in the questions each can answer: search handles “find the Meridian contract”; the layer handles “what do we owe Meridian and why?”
• It is not a wiki or a knowledge base. Wikis rot because humans must maintain them, and humans have jobs. The layer is written to by the tools doing the work, so it stays current as a side effect of the work.
• It is not a vector database. If your team has technical folks, they may ask. Search by meaning is a useful index inside a context layer, but “throw everything in and search by similarity” produces a bag of fragments, not a picture of a business. The structure, typed memories, links, history, is the product; similarity search is plumbing.
• It is not tied to one AI vendor. Models will leapfrog each other every year. The layer is where accumulated knowledge lives, so you can swap tools freely and lose nothing. Why that may matter more than model choice is the subject of One brain, many agents.

The questions owners actually ask

“Where does it run?” Context Engine is a hosted service at app.contextengine.com. Each engine, which is what we call one memory, gets its own database, its own search index and its own Worker, so isolation is physical rather than a filter on a shared table. What you save is never used to train models, and you can export the whole engine as Markdown and JSON that opens without us. Piece five takes the ownership question head on.

“Can it act on its own? Send things?” No, and this is a design principle rather than a missing feature. The layer is a library, not an actor: it holds memory and answers questions. It does not run your agents and it does not reach into your other apps. Your AI tools, which you authorize, do the acting.

“What does my team have to change?” Very little, and that is the honest adoption test. Each person connects their AI tool once and gets used to saying “save that” at the end of work that matters. Every so often someone looks at Memories, answers the duplicate questions on Home, and corrects what is wrong, which is what keeps the memory trustworthy. Everything else is subtraction: context that is already there instead of being pasted in.

“How is this different from just buying more AI tools?” Tools are surfaces; the layer is what they share. Without it, every new tool adds another silo. With it, every new tool arrives already knowing the company, and every tool you retire leaves its knowledge behind.

Build this yourself

1. Create an engine and give it a purpose: a few sentences on what it is for and what does not belong. AI tools read the purpose to decide what to save there.
2. Connect two AI tools you actually use, say Claude and ChatGPT, each with Can read and save. For Claude Code it is one command: claude mcp add --scope user --transport http ce-clients https://app.contextengine.com/mcp/c/<id>, with the address the app gives you.
3. Have one tool save a decision and its reason. Then ask the other tool what was decided, and check Memories to see both the decision and who saved it.

Start free, or see plans.

Why “layer” is the right word

Because it is infrastructure. Your business already accepted this pattern once: every serious company runs on a database, and no one considers that optional. The context layer is the same move for the AI era. The tools on top will change constantly. The memory underneath should be stable, structured and, crucially, yours. Next in this track: how the layer actually structures a business, and further on, the question that follows from concentrating all this in one place: who should own it.