Every major AI release focuses on the same metrics.
- More tokens
- Better benchmarks
- Faster inference
- Lower latency
Those are important.
But they’re not the bottleneck most teams experience after months of AI-assisted development.
The bottleneck is this:
Nobody remembers why anything was built.
⸻
Software development has entered a strange phase.
Writing code is becoming easier.
Understanding existing code is becoming harder.
Ironically, AI accelerates both trends.
It produces software faster than teams can absorb the decisions behind it.
⸻
Think about a mature repository.
It doesn’t just contain source code.
It contains years of trade-offs.
Rejected designs.
Temporary compromises.
Performance fixes.
Security decisions.
Business requirements.
These pieces of knowledge are often more valuable than the implementation itself.
Yet they’re the first things to disappear.
⸻
Perhaps repositories need a second layer.
One layer stores code.
Another stores understanding.
One records implementation.
The other records intent.
When AI tools can access both, they stop acting like code generators and start behaving like long-term engineering collaborators.
That feels like a much more interesting direction than simply making context windows bigger.
https://www.contorium.dev/
https://github.com/ContoriumLabs/contorium
Top comments (0)