Context is the umbrella abstraction for long-running AI agents. It includes the state needed to continue an execution, the workspace where agents do their work, durable memory carried across sessions, and knowledge synthesized from that memory.
In short:
Context = Execution + Workspace + Memory + Knowledge
Context Service aims to make these layers durable, accessible, and manageable across agents without requiring every agent system to build its own storage and lifecycle mechanisms.
Execution context is the active state of an agent: harness state, prompts, tool history, and conversations. It lets an interrupted or restarted agent continue its work.
Workspace context is where work happens: Git repositories, sandbox filesystems, source files, generated artifacts, and other execution assets.
Memory is durable experience carried beyond one execution or session. It may include observations, decisions, Markdown documents, reports, PDFs, and other structured artifacts.
Knowledge is the understanding distilled from memory into reusable forms such as living wikis, documentation, reports, and knowledge bases.
These layers are related but distinct. A workspace is not automatically memory, and a collection of memories is not automatically knowledge.
Agent memory is often treated as a capability embedded separately in each agent system. The broader opportunity is to treat memory—and eventually context—as shared infrastructure.
A Context Service could provide common capabilities for storing, organizing, retrieving, sharing, governing, synchronizing, and evolving context across long-running and collaborative agent systems.
The library is a useful analogy:
- memories are the books;
- the Context Service is the library system that manages cataloging, access, provenance, versions, and lifecycle;
- the knowledge layer continuously studies the collection and produces higher-level understanding.
Memory captures what happened. Knowledge captures what has been learned.
This model creates opportunities for:
- shared memory and workspaces across agents;
- promotion of useful information from transient execution into durable memory;
- collaborative memory management with clear ownership and provenance;
- living knowledge bases that agents continuously refine;
- context that survives individual agents, sessions, and compute instances;
- a common context layer usable by different agent runtimes and platforms.
The vision is intentionally broader than the first implementation.
This repository begins with one concrete problem: dynamically creating sandbox compute with durable workspace storage for serverless-harness. That small slice provides a place to learn what the service must own, which abstractions are real, and where the boundaries belong.
Broader memory and knowledge capabilities will be added only as working use cases make them clear.