September 2026 · Northforged Labs
One context. Every AI.
The model wars will produce several winners. The context layer only needs one — a human-owned compiler so the next AI receives not your history, but the smallest package of truth it needs to continue your work.
Press C to open the compiler. Prepared by Northforged Labs.
01 · The problem
The most expensive thing in AI isn’t tokens. It’s the human, re-explaining.
Strategy in ChatGPT. Writing in Claude. Code in Cursor. Images and research each in their own tool. Each is a sealed room. The only thing that moves between rooms is whatever you carry in your head and type again.
Today
- “Let me explain my business again.”
- Copy, paste, trim, re-paste between models.
- Every AI keeps its own private, unexportable picture of you.
- Big windows filled with stale history. The model guesses which decision is current.
- Switching tools has a cost, so you stop switching — even when a better tool exists.
With ContextOS
- “Create the logo for the business I just designed.”
- Each model pulls the context it needs, when it needs it.
- The project is the source of truth. Every AI reads from it.
- A compiled handoff: ~1,000 tokens of what is true now, not 40,000 of what was said.
- Switching is free, so you use the best tool for each job.
75–170 hours / year
Six to eight context switches a day, three to five minutes each. For a founder or senior engineer at $100/hour, that is $7,500–$17,000 a year spent telling machines what they already told other machines.
02 · The product
Not a memory. A compiler.
ContextOS ingests work from sources you approve, extracts durable claims, maintains a project graph, and compiles — for a given task, target and token budget — the smallest high-value package that lets the next AI continue without being told anything twice.
-
01
Capture
Browser extension and desktop app for ChatGPT, Claude, Cursor, VS Code. Connectors for files and repos. The labs’ own memory exports as the bootstrap. Every source is opt-in. Nothing is captured silently.
-
02
Extract
Raw activity becomes typed claims, not summaries: decisions, rejected alternatives and why, constraints, preferences, entities, artifacts, open questions, next actions. Each claim carries provenance and a validity window.
-
03
Graph
People ↔ projects ↔ decisions ↔ artifacts ↔ tasks ↔ questions. The system of record for the work — what today lives in fourteen tabs, three model memories and one head.
-
04
Compile
Input: task, target, budget, permissions. Output: the minimal package, in the dialect the target performs best with. Usage and corrections flow back into ranking. Every handoff improves the next.
Chat history is a log. ContextOS is a ledger. A log records that you chose Stripe in February and Paddle in June. A ledger knows that Paddle is current, why, since when — and that Stripe is history unless the task needs history.
03 · The five-minute demo
Spend an hour designing a business in ChatGPT. Open Claude. Nobody explained anything twice.
The highest-value tokens are the rejections. Summaries drop them first. ContextOS treats a rejection as a first-class object with a reason attached, and ranks it near the top of any creative or strategic handoff.
compile(task, target, budget)
04 · The graph
Why this is not RAG over your chat history.
Retrieval returns passages that resemble the question. The compiler returns the state of the work: what is true now, what was decided, what was ruled out, what is next.
Project
Decision
Rejection
Artifact
Source
- Bi-temporal truth. Claims carry valid-from and valid-to. Superseded decisions stay for provenance and leave the handoff unless the task asks for history.
- Decision-aware, not fact-aware. Rejections, constraints and objectives are typed objects, ranked by task — not paragraphs found by similarity.
- Task-conditioned selection. “Write the pitch” and “fix the auth bug” draw from one graph and share almost nothing.
- Target-aware formatting. A system-prompt block for one model, a markdown rules file for another, a structured tool result over MCP for a third.
- Closed-loop weighting. Every handoff is an experiment. The data asset is not the raw chats. It is the learned map of what matters for which kind of work.
05 · Why now, and why us
Four things happened in twelve months. None of them existed for this idea in 2024.
-
The labs admitted the problem, then answered it with a one-way door.
Claude, Gemini and ChatGPT all shipped memory export or import within three weeks in March 2026. Import into me, never sync between us. Those exports are our bootstrap.
-
Memory became a first-order lab priority — and stayed single-app.
ChatGPT’s rebuilt memory lifted factual recall from 41.5% to 82.8%. It belongs to one application, and will always be built to keep you inside it. The neutral layer cannot come from a lab.
-
Distribution got solved before we started.
MCP was donated to the Linux Foundation with Amazon, Google, Microsoft and OpenAI as founding members. A context layer that ships as an MCP server is readable by every major client on day one. No permission required.
-
Bigger windows made compilation more valuable, not less.
Context Rot showed reliability falling with input length across 18 models. The scarce resource is relevance. Relevance is a compiler problem.
The landscape, by who owns the memory and how far it travels.
Lab memory
ChatGPT · Claude · Gemini. Excellent. Single-app. Built to retain.
Developer memory
Mem0 · Zep · Letta. Sold to the developer. Memory belongs to that app.
Exports & formats
Portable, once. One-way. No compiler, no continuity, no product.
ContextOS
Human-owned. Cross-app, continuous. Compiled per task. The quadrant every incumbent is paid to avoid.
$2.59T
worldwide AI spend in 2026 · $3.49T forecast for 2027. We don’t need that whole market. We need the people already paying for three tools and feeling the context tax.
Gartner · May 2026
$15–20
per month for the layer that makes the $100–200 you already spend actually work together.
~15M
power users worldwide already spending $1,200–$2,400 a year on AI. We are not asking them to start paying.
$24B
of the $40B consumer market sits in people paying $100+/month. That is the beachhead.
06 · Access
This page exists to be argued with.
If the idea is wrong, better to hear it from you than from the market. If it is right, you already know why — you have been paying the context tax longer than almost anyone.
The model wars will produce several winners.
The context layer only needs one.