Memory Orchestrator

Implement a layered memory system that remembers the right things, loads the least necessary context, and improves over time through reflection.

Core Model

Represent memory in five layers:

1. Session state — current task, active domains, active objects, constraints
2. Daily raw log — important interactions and memory-worthy events from the day
3. Topic index — summary cards for broad domains. Seed topics like technology/career/investing/research/life/meta are only defaults, not a fixed ontology.
4. Object store — durable objects such as papers, concepts, frameworks, decisions, preferences, open questions, notes, people, projects, or any future type that proves useful
5. Reflection output — compressed updates that merge duplicates, raise abstraction, and add cross-topic links

Do not use one giant memory file as the primary operating model.

Default Directory Layout

Use this layout inside the workspace unless the user specifies another storage backend. Treat it as a readable default, not a rigid schema: