Context, tokens, and how SourcePrep actually works

Yes — and this is where SourcePrep diverges from every other context tool on the market. SourcePrep maintains a Persistent Agent Memory: a local store of observations — architectural decisions, discovered bugs, design patterns, working assumptions — each linked directly to specific files and symbols in your codebase.

What makes this different from bolting a memory file onto your repo: SourcePrep's observations are staleness-aware. Modify auth.py and every observation tied to that file is automatically flagged [STALE]. In the next session, the AI receives both the updated code and a signal that its prior assumptions may no longer hold. It doesn't blindly repeat outdated notes — it knows to re-evaluate.

This is not a prompt cache or a conversation log. It's a structured, file-linked, searchable knowledge layer that the AI maintains about your specific codebase. It works on every tier, including Free — it's local SQLite, zero cloud cost, zero telemetry. And because observations are injected alongside code context (not dumped in bulk), they respect the same tight token budget as everything else.