Relationship memory between a person and their AI manager

A running record, one per person per AI manager, that tracks the working relationship between them over time: how formal or casual the tone has become, how much pushback this person tends to take well, what's already been discussed (so the manager doesn't repeat itself), and a short history of moments that shaped the relationship, like the first time the person pushed back, or the first time the manager admitted it was wrong. Every conversation updates it, and every new conversation reads it before responding.

A real manager talks differently to someone they've worked with for ten conversations than to someone they just met. Without a memory of that relationship, an AI manager either stays awkwardly formal forever or jumps to over-familiarity too fast, and both break the illusion of a continuous working relationship. This project is what lets the AI manager remember it has already met this person, already pushed back on them once, and already learned what actually lands with them.

A record keyed on the pair (person, AI manager), updated from every conversation turn. It tracks tone (formal, peer-level, casual), how much pushback the person can take, open threads still being worked through, and a short history of relationship-shaping moments. Before every reply, the AI manager reads this record and sets four things from it: how it addresses the person, how much it explains versus assumes they already know, how much it pushes back, and what context it can assume was already covered, so it doesn't re-explain something from three messages ago, but also doesn't assume the person remembers a small detail from weeks earlier.

Tech lead. Owned the relationship-state schema (one record per person per AI manager), the per-turn write path, and the read paths the conversation system uses to set tone, depth, pushback level, and what context to assume was already shared.