Large language models continue to improve at writing code.
But one problem keeps slowing developers down:
Context Drift.
Every new AI conversation gradually loses awareness of the project.
The larger the repository becomes, the worse the problem gets.
Typical symptoms include:
- AI suggesting deleted architectures
- forgetting previous design decisions
- generating duplicate implementations
- inconsistent naming
- conflicting refactors
These aren’t failures of reasoning.
They’re failures of project continuity.
A Different Direction
Instead of trying to increase chat memory indefinitely, I’ve been experimenting with another idea:
Move project intelligence into the repository itself.
Rather than asking AI to remember everything, let the project expose structured knowledge that any AI can inspect.
This includes information such as:
- project state
- historical evolution
- architectural relationships
- decision provenance
- developer intent
The result is that AI spends less time reconstructing context and more time solving actual engineering problems.
That’s the direction I’m exploring with Contorium.
It’s still evolving, but the long-term goal is straightforward:
Projects should have persistent intelligence—not just persistent code.
GitHub:
https://github.com/ContoriumLabs/contorium
I’d love to hear how other developers are approaching context continuity in AI workflows.
Top comments (0)