AI coding agents already have access to a lot of context.
They can index repositories, search files, retrieve documentation, inspect Git history, use MCP tools, and maintain various forms of session memory.
But while building Contorium, I kept running into a different problem:
Repository context is not the same thing as project understanding.
The missing layer
A repository can tell an agent:
src/
auth/
api/
database/
services/
It can tell the agent what functions exist and how modules depend on each other.
But it may not tell the agent:
Why was this architecture chosen?
Why was another approach rejected?
Which constraints are intentional?
Which decisions are still valid?
What changed after the original design?
What is the current project state?
This information is often distributed across commits, conversations, issues, documentation, and — most importantly — the developer’s own reasoning.
That is the layer I’m experimenting with.
Project Intelligence
Contorium defines this layer as Project Intelligence.
At a high level:
Intent
↓
Decisions
↓
Reasoning
↓
Implementation
↓
Outcomes
↓
Evolution
↓
Current State
The important part isn’t just storing these objects.
It’s connecting them.
For example:
Decision
│
├── motivated by → Constraint
│
├── implemented by → Module
│
├── replaced by → New Decision
│
└── resulted in → Outcome
This creates a project-level knowledge structure rather than another flat memory store.
Where Contorium sits
The architecture is intentionally separated from the AI coding tool:
┌─────────────────────────────────────────┐
│ AI Tools │
│ │
│ Cursor · Claude Code · Codex · VS Code │
└───────────────────┬─────────────────────┘
│
▼
CIL — Cognitive Interaction Layer
│
▼
Project Intelligence
│
▼
.contora/
│
▼
Your Project
CIL is the interaction layer.
Project Intelligence is the persistent project-level state.
.contora/ keeps that intelligence associated with the project itself.
This separation is important because the AI runtime shouldn’t own the project’s accumulated understanding.
Why not just use Git?
Git already contains history.
But Git answers questions like:
What changed?
Project Intelligence tries to answer:
Why did it change?
And then:
What happened afterward?
And eventually:
Does the original reasoning still apply?
Git is an essential source of project history.
It isn’t the entire semantic model of the project’s evolution.
Why not just use RAG?
RAG is very good at retrieval.
If the question is:
Find the documentation about authentication.
retrieval is exactly what we need.
But a project intelligence system may need to connect:
Authentication
│
├── Decision
│ └── Why JWT was rejected
│
├── Constraint
│ └── Existing mobile clients
│
├── Module
│ └── auth-service
│
└── Outcome
└── Migration completed
The interesting part isn’t just retrieving the relevant documents.
It’s understanding the relationships between them.
Why this matters for AI agents
Imagine switching from one AI coding tool to another.
The new agent can read the repository.
It can inspect the current code.
But without project intelligence, it may still need to reconstruct the project’s history and reasoning.
With a persistent project layer, the workflow becomes:
New Agent
↓
Read Project Intelligence
↓
Understand Decisions
↓
Understand Constraints
↓
Understand Current State
↓
Continue Development
The model changes.
The tool changes.
The project intelligence remains.
Contorium today
Contorium is currently being built around several pieces:
- project history
- structured project state
- decision and reasoning tracking
- project graph
- AI-ready handoff
- MCP integration
- IDE integration
- local .contora/ storage
The project is open source, and I’m deliberately building it in public.
I’m not claiming that Project Intelligence is a solved problem.
I’m trying to find the right abstraction for it.
The bigger question
AI coding is moving from:
Human → Prompt → Code
toward:
Human
↓
Project
↓
Multiple AI Agents
↓
Continuous Evolution
Once multiple agents work on the same project over long periods, something needs to preserve the project’s accumulated understanding.
That’s the problem I’m interested in.
Not just AI memory.
Not just more context.
Project Intelligence.
That’s the direction behind Contorium.
https://www.contorium.dev/
https://github.com/ContoriumLabs/contorium
Top comments (0)