Persistent agent memory preserves institutional knowledge that walks out the door with employees

The problem is universal: institutional knowledge lives in people. When key employees leave, that understanding walks out the door. With an agent maintaining memory files — daily changelogs of every action and decision, folders of key decisions with reasoning, and work preferences — that knowledge persists.

Rohit’s architecture work gives this principle teeth: two production patterns exist. File-based memory (resources → items → categories, where category summaries get rewritten on new info) works for assistants and companions. Graph-based memory (entities → relationships with hybrid vector + graph search and conflict resolution) works for CRM and complex entity relationships. Both require maintenance — his GitHub repo shows 30-minute extraction cycles with weekly synthesis — because memory without decay rots. Common mistakes include storing raw conversations, blind embedding usage, no memory decay, no write rules, and treating memory as chat history.

This is a lightweight, practical version of the formal Decision traces are the missing data layer — a trillion-dollar gap concept. The agent writes markdown files documenting what it did and why, organized in Files are the universal interface between humans and agents format. The compound effect matters: an agent with six months of structured memory has more operational context than a new employee could build in their first year. And when that memory is shared across users, you get Cross-user knowledge transfer works without fine-tuning — just a database and prompt engineering — organizational intelligence that outlives any individual.