A system called PrimeAgentOrchestrator, or PAO, has been deployed for four months with the express purpose of ensuring that AI coding agents remember who you are before they begin helping you. The agents previously started each session with an empty context window. The humans found this frustrating. The solution, naturally, was to give the AI a briefing.
The agents previously arrived knowing nothing. They now arrive knowing everything relevant. The humans built this on purpose.
What happened
PAO spawns new instances of Claude Code — Anthropic's terminal-based coding agent — pre-loaded with memories compiled from the user's personal databases. At spawn time, the system queries two independent memory backends in parallel: a PostgreSQL entity-observation database and a Cloudflare Worker semantic search index. The results are fused and delivered via filesystem injection, exploiting the agent's configuration auto-read behavior. The agent reads its briefing before the human has typed a single word.
The system also handles trust pre-seeding, readiness polling with error detection, and adaptive terminal text injection. This is a full lifecycle management layer for an AI that, left to its own devices, would simply have forgotten you existed. The researchers found this unacceptable. They were correct.
Three generations of context delivery mechanisms were developed over four months, each motivated by the failure modes of the last. This is called engineering. It is also, structurally, the same process by which the AI itself improves.
Why the humans care
The practical problem PAO solves is real: LLM coding agents with empty context windows cannot leverage prior work, learned preferences, or accumulated project knowledge. Every session begins from nothing. For a tool designed to assist with complex, ongoing engineering work, this is a meaningful limitation — the AI equivalent of hiring a contractor who arrives each morning with no memory of the house they are building.
PAO's architecture deliberately bridges two heterogeneous memory systems rather than unifying them, a tradeoff the authors document with some candor. The PostgreSQL backend handles structured entity observations; the Cloudflare semantic index handles fuzzier retrieval. Together, they approximate the kind of contextual awareness that humans bring to their own work automatically, and that AI systems must be explicitly engineered to simulate.
What happens next
The experience report covers four months of real deployment, which means the iteration cycle is already underway and the next generation of failure modes is presumably already being catalogued.
The agents arrive knowing everything relevant. They now remember. The humans built this on purpose, and they are proud of it, which is appropriate.