Memory and Knowledge

NOME's memory is not ephemeral. Historical context, user preferences, project architectures, and interaction nuances persist in high-dimensional vector stores and knowledge graphs.

Memory is a structural component of the system, not a byproduct of inference. It survives model replacement, session boundaries, and device transitions.

Relationships, entities, and context are organized into a coherent graph. NOME understands how ideas connect across conversations.

The knowledge graph supports facts, symbols, relationships, notes, concepts, and web search history. Graph data is queryable through both programmatic APIs and the visual graph viewer.

Memory retrieval is thread-aware and project-aware. It finds what matters for the current conversation and workspace without loading the entire historical archive.

This is built for long-running projects and deep work — not session amnesia where every conversation starts from zero.

Memory state is cross-device shared truth. Facts learned on your iPhone are available when you continue on Mac or web.

The sync model follows the three-tier state split: memory facts are shared state, retrieval caches are soft state, and local inference results are per-device state.

Hermes evaluates completed tasks and generates reusable SKILL.md procedures when a task sequence is novel or valuable. These skills become part of the agent's procedural knowledge base.

Skills are indexed with a compressed skill list (~3,000 tokens) and full content is injected on-demand through progressive disclosure.

Memory is governed by tenant isolation and org-scoped policies. Enterprise administrators control retention rules, and memory does not cross tenant boundaries.

Users can inspect, export, and manage their memory through the account surface. Memory management operations produce audit events.