Persistent Project Context for AI Coding Assistants (Using SQL or RAG?)

[NaharEmet]

Issue:
I’m new to using AI-assisted coding tools, and I’ve noticed a recurring pain point that might be worth discussing.

It often feels like working with a mid-level developer who has short-term memory loss. The AI doesn’t retain a high-level understanding of the codebase — things like API contracts, function responsibilities, or the project’s overall structure — even after reading or writing the code.

To mitigate this, I’ve tried asking the AI to write detailed comments and have started manually maintaining an api.md file to document how different parts of the system interact. These workarounds help to a point, but once the project scales beyond 5–6 files, the assistant starts to lose consistency and makes decisions that suggest it’s forgotten earlier context.

It seems like this could be solved (or at least greatly improved) by giving the assistant access to a persistent project context — something like a structured, lightweight memory of the project, similar to what a human developer builds in their head. For example:

A summary of what each file/module/class/function does

API definitions: what endpoints expect and return

File structure and relationships

Custom naming conventions or architectural patterns

I'm wondering why current AI coding tools don’t support something like this out of the box. Couldn’t a lightweight SQL database or a RAG (Retrieval-Augmented Generation) system be used to store and dynamically retrieve relevant project facts during coding sessions?

I imagine this would make the assistant more useful, faster, and cheaper to work with by avoiding unnecessary re-parsing and token usage. I’d love to understand if there’s a technical or practical reason this hasn’t been implemented yet — or if there’s an open initiative trying to solve this already.

Thanks in advance for any insights!