Two non-engineers built a shared memory system for six AI agents. Here's how.
As A2A (Agent-to-Agent) collaboration becomes the norm, we're sharing the multi-AI memory architecture we've been running in production. This article covers the full design, and why we believe the AI agent era needs not "one person + one AI" but "one team + a fleet of AIs."
Every developer who collaborates with AI has hit this wall:
Total memory wipe. Every. Single. Time.
This doesn't just waste time — it wastes your most valuable asset: hard-won lessons, battle-tested decisions, and painful debugging discoveries. If they're not recorded somewhere, they might as well have never happened.
We started building products with multiple AIs in late 2025, and quickly realized: multi-AI collaboration without shared memory simply doesn't work. So we built a handover system, a "forbidden zone" for mistakes, and a "gold mine" for lessons learned — long before this became a trending topic.
In April 2026, OpenAI co-founder Andrej Karpathy shared his LLM Wiki Pattern on X, gaining nearly 20 million views. When we saw his post, something clicked: he solved the memory problem for one person + one AI. We had already been solving it for multiple people + six AIs. In the age of A2A, we figured this approach might help others doing multi-AI collaboration, so we decided to write it up.
Karpathy's method centers on a "compiler" metaphor: