I Designed a RAG Variant for Multi-Agent Simulations. Here's the Design and the Honest Tradeoffs.

Standard RAG is great for static knowledge bases. Embed documents, embed a query, return top-k by cosine similarity. That works.

But put RAG inside a running civilization where 40 citizens have memories, councils deliberate on crises, and past decisions ripple into future ones, and similarity alone breaks down fast.

The problem is simple: cosine similarity doesn't know that last month's drought caused today's food riot. It doesn't know that the council that voted against emergency grain reserves three weeks ago is directly responsible for the current famine. It retrieves memories that sound like the crisis, not memories that led to it.

That gap is what I wanted to close. This post explains the retrieval design I built for .

If you're building multi-agent simulations or agentic systems where decisions have downstream effects, the core idea is worth stealing: semantic similarity and causal relevance are not the same thing, and for an agent making decisions under pressure, the difference matters.

If you're working on agent memory or causal retrieval, I'd genuinely like to hear how you're handling it - reply here or find me on .