An Engineer Spent Four Months Trying to Give His AI Coder a Memory

Every new Claude Code session starts with an empty context window. The agent has no memory of the codebase, the user's preferences, or yesterday's debugging. An arXiv experience report describes one engineer's four-month attempt to fix that with a system he calls PrimeAgentOrchestrator.

The author ran the system on his own setup from December 2025 through March 2026. It pre-loads each spawned Claude Code instance with a compiled briefing drawn from the user's existing personal databases. At spawn time it queries two independently operated memory backends in parallel: a PostgreSQL entity-observation database and a Cloudflare Worker semantic search index. It fuses the results with backend-specific retrieval strategies, then delivers the briefing by writing files the host agent already auto-reads on startup.

Three generations of context-delivery mechanisms are documented in the paper, each redesigned after the previous one failed in practice. The author frames the design as a deliberate tradeoff: bridge heterogeneous memory systems rather than build one unified store. The system is single-user, self-reported, and not independently benchmarked. It is a practitioner's working answer to a real problem, not a path to general agent infrastructure.