Graph-native persistent memory for constitutional AI systems.
ODEI Memory documents the memory architecture behind ODEI's world model:
- typed graph memory instead of flat conversation logs
- structural retrieval instead of similarity-only recall
- guarded writes instead of direct, unaudited mutation
- provenance, temporal context, and auditability as first-class primitives
Most agent stacks still treat memory as a cache or transcript problem. ODEI treats memory as infrastructure.
That means memory must support:
- durable identity and long-horizon context
- explicit relationships between goals, decisions, tasks, and evidence
- constrained mutation paths
- retrieval that respects structure, time, and policy
- Graph-native: relationships are part of the memory model, not metadata bolted on later
- Local-first: the world model should remain portable, inspectable, and user-owned
- Constitutional: writes are validated against system invariants before they become state
- Auditable: important decisions and mutations remain reconstructable across time
- Budget-aware: retrieval is shaped for real agent runtimes, not idealized demos
ODEI memory is built around a constitutional world model that separates:
- identity and principles
- direction and goals
- strategic plans
- tactical systems
- execution state
- observations, signals, and evidence
This allows agents to answer structural questions such as:
- what decisions led to this task
- which goals are blocked by this dependency
- what changed since the last execution cycle
- which facts are observations versus inferred beliefs
- odei.ai is the research surface
- api.odei.ai is the production surface
- this repository focuses on the memory layer that makes both coherent
| Repository | Role |
|---|---|
| web | Public product, docs, and API runtime |
| research | Papers on constitutional world models and agent memory |
| mcp-odei | MCP server for governed retrieval and guardrail checks |
| examples | Integration examples for API and MCP workflows |
MIT