Agent memory
Agent memory is the context an AI agent retains and reuses beyond a single interaction. In nOS it is modeled in three layers: working memory that is private to the agent, shared memory where multiple agents coordinate, and verifiable memory where context carries provenance any party can check.
Working memory
Private to the agent that creates it. Drafts, intermediate reasoning, and early context live here, inside your own infrastructure, before anything is shared. Most consumer assistant memory never leaves this layer.
Shared memory
Where several agents draft, refine, and reuse context together. This is where collective intelligence appears: context created once gains value as more agents contribute to and rely on it.
Verifiable memory
Where source, lineage, and provenance remain attached, so context can be reused with a history that others can check. This is the layer that lets memory cross an organizational boundary without requiring trust in whoever produced it.
In practice
A support agent drafts an analysis of a recurring fault in working memory. Once reviewed it moves to shared memory, where a field-service agent and a warranty agent both build on it. When the warranty agent uses it to justify a claim decision, the conclusion is written to verifiable memory with its lineage, so the manufacturer can later show the claim rested on a traceable finding rather than an assertion.
Common questions about agent memory.
Why not just give every agent a bigger context window?
Does memory sharing create a privacy problem?
Which layer do decision traces live in?
Read next.
Power up your business with the OriginTrail DKG.
Create Knowledge Assets, build shared context graphs, enshrine decision traces. Open-source infrastructure. Community support. Zero cost.
The full Network Operating System.
Custom agent frameworks (Hermes, OpenClaw, LangChain, Claude). Custom data pipelines. Custom integrations with your enterprise systems. Dedicated infrastructure and support.