Hi HN,

I’ve been experimenting with building LLM agents and kept running into the same problem: managing context gets messy quickly. Prompts grow, reasoning steps get lost, and it's hard to reproduce or roll back agent behavior.

So I built Contexa, a small open-source project that treats LLM context a bit like Git — with ideas like snapshots, branching, and history for context.

The goal is to make it easier to:

track how context evolves

experiment with different reasoning paths

reproduce agent behavior

It’s still early, but I’d love feedback from people building LLM tools.

Repo: https://github.com/swadhinbiswas/contexa Paper: https://arxiv.org/abs/2508.00031