A memory layer for AI agents

The bet

AI coding tools forget. They generate code without context, repeat decisions you’ve already overruled, lose track of what your team agreed last week. A memory layer fixes that — but only if it survives concurrent users without corrupting itself. The first prototype didn’t. This case study is what it took to make it.

What it is

A memory layer that augments AI coding harnesses with persistent, retrievable context. Multi-agent safe, governance-aware, fast enough not to break flow. Built for personal use first; now adapted to run the knowledge store workflow at a tech-driven bank.

Decisions that mattered

• SQLite → Postgres for concurrency. File-locking doesn’t survive multi-agent writes.
• Graph-over-RAG for reasoning. Embeddings find candidates; graph edges reconstruct context.
• Session-set admin guard. Authority changes require a session variable + trigger — not just a flag flip. Governance in miniature.
• 97% size reduction from prototype to production. Cleaner data model, smaller working set, no recall loss.

What’s open

How AI agents share memory without stepping on each other under sustained concurrent load. Still under active research.

In production at a tech-driven bank. Reach out if your org needs the same.