Mnema Mnema
Join the waitlist
← All docs Flows

Flows

Flows are how Mnema turns a workspace of docs into structured context for an AI agent. Here's what they are, how Claude calls them, and when to reach for one.

  1. 1

    What a flow is

    A flow is a named sequence of content steps. Each step has an instruction and references one or more docs in your workspace. Flows are the unit of context that you deliver to Claude — not raw docs, not entire workspaces, but specific ordered sequences with your authorial intent baked in.

  2. 2

    How Claude calls a flow

    Claude (or any MCP client) discovers your flows by calling list_flows. It then walks the flow by calling get_flow_step(flow_id, step_number) for each step in order. Each call returns one step's instruction plus the rendered content. The sequence is explicit — Claude doesn't decide order, you do.

  3. 3

    When to use a flow vs. a single doc

    Single docs are good when Claude needs a fact or a specific piece of content. Flows are good when Claude needs to be walked through something — a codebase onboarding, a workflow, a structured analysis. If you'd brief a teammate over five minutes with "first read this, then this, then this" — that's a flow.

  4. 4

    Capture nodes — Claude writes back

    Most nodes give Claude something to read. A capture node asks Claude to produce something and write it back as a new doc in your workspace — so a later step (or a later flow) can read it. You author the node with a title hint and an instruction ("research this vendor and write your findings"); at walk time Claude generates the content and calls submit_flow_capture. By default the write is gated: it surfaces as a proposal you approve before the doc is created — the same approval you already use for any doc write. A capture node can be marked autonomous at authoring time, which lets Claude write the doc directly with no review step. That toggle carries a warning in the editor for a reason — turn it on only when you trust the flow to write unattended.

  5. 5

    The flow editor (coming next release)

    The visual flow editor is in active development. Drag docs into a sequence, annotate each step with an instruction, save the flow. It's exposed to Claude automatically through MCP — no extra wiring needed.

Today vs. next release

The MCP server already advertises list_flows and get_flow_step as tools. They currently return an empty preview response — the flow editor fills them in. The contract is stable now so Claude conversations referencing flows will keep working when the editor ships.

submit_flow_capture reference

Called by Claude when a walk reaches a capture node. Arguments: flow_slug, node_id, title, markdown, and an optional target_folder_id. Whether the write is gated or autonomous is decided by the node, not the caller — Claude cannot escalate its own permissions. It needs a workspace-write token. Autonomous writes are rate-limited and capped at 100k characters per capture, and every one is recorded in the workspace audit log with its originating flow and node. (For now that provenance lives in the audit log rather than on the doc itself.)