I've been designing a memory system for Isomux.
In Isomux, agents are organized hierarchically: Office > Room > Agent
Features are shaped around this mental model.
For example, each layer has its own user-defined system prompt, which is inserted at the beginning of each session.
I want memories to follow the same hierarchy: they can be for a specific agent, a whole room, or the whole office.
Memory creation: a hook that runs after every turn. It extracts facts to remember, chooses their scope (agent/room/office), and stores them.
Memory retrieval: before every assistant turn, we'd pull relevant memories for this agent, the room they're in, and the office-wide context.
I experimented with an existing memory solution, mem0, but it doesn't fit this hierarchical memory shape. I'll probably roll my own following their open-source library.
