Skip to main content

ICM Workspaces

Coming soon

ICM workspaces are not generally available yet. This page describes the method and where MemoryRelay is going. There is no public API to call yet, and nothing here is a specification. To hear when early access opens, email hello@memoryrelay.ai.

MemoryRelay is becoming an ICM service: a hosted home for the context your agents work from, organised with the Interpretable Context Methodology.

The method​

ICM keeps what an agent needs to know in a small workspace of plain files. Instead of searching everything and hoping the right pieces come back, the agent follows the workspace's structure to the stage it is working on.

  • Folders carry sequencing. Work is split into ordered stages. Each stage is a folder with a contract: what it takes in, what it produces, and who checks it.
  • Hierarchy scopes context. The agent walks from one short entry file to the stage it needs, and loads only what that stage lists.
  • Plain files carry state. Context, decisions and outputs are ordinary files that people can read, review and keep in git, and that an agent can walk with no service running.

An illustrative workspace (a pipeline; workspaces take several shapes):

release-notes/
├── CLAUDE.md          # entry: where to go
├── CONTEXT.md         # always-true context
└── stages/
    ├── 01-collect/
    │   ├── CONTEXT.md # stage contract
    │   └── output/
    ├── 02-draft/
    └── 03-review/

What MemoryRelay is building​

The method works with nothing but a folder. MemoryRelay will host, version and deliver those workspaces:

  • Versioned releases: publish a workspace as an immutable release, so every agent run knows exactly which context it was given.
  • Stage-scoped delivery: hand an agent the context for one stage, within a budget, instead of a search result it has to trust.
  • Human review, recorded: every stage ends at a human check, tied to the output that was actually reviewed.
  • Memory as evidence: your MemoryRelay memories feed in as evidence with their provenance, and are never silently promoted to instructions.

How it relates to the memory API​

Everything else in these docs, including memories, semantic search, entities and the MCP server, is the memory layer. It is available today and is what ICM workspaces will draw on. You don't have to wait for ICM to start storing what your agents learn.

Attribution​

ICM is described in Interpretable Context Methodology by Jake Van Clief and David McDermott (arXiv:2603.16021v2). MemoryRelay follows the ICM Architect reference (MIT licence). MemoryRelay's service is an independent implementation and is not endorsed by the authors.