mwe — the memory-first assistant template

The default consumer agent of an mwe-mcp memory server, as a nanoclaw agent template (Agent Plugins 1.0.0).

Stamp it into a nanoclaw fork:

ncl groups create --template mwe --name mwe --new

--name is yours to choose — it becomes the group folder and how the agent shows up in ncl. It is not what the agent answers to: the bridge stops nanoclaw injecting a configured name, so the agent's name is a fact of the memory (WHO YOU ARE) that anybody it serves can give it, in chat. The product imposes no character either: it is the household's assistant, not a brand.

What it gives the agent

The persona in ai.nanoco.nanoclaw/context/instructions.md is prepended to the agent's system prompt on every spawn, under every provider. It teaches one thing: mwe-mcp is this agent's only memory, capture and recall are mechanical, and there is no file on disk to keep a second copy in.

What it does not do on its own

The template is the persona. The mechanics — one ingest per turn, the recall block, per-sender attribution, the reverse channel — are the add-mwe-memory skill, which patches the fork. A group stamped from this template without that skill installed is a normal nanoclaw agent that has been told about a memory it cannot reach.

Both arrive from the same bridge, and either order works:

• Template first is what nanoclaw's setup wizard does — it stamps the first agent before you have a skill to apply. Applying the skill afterwards clears the memory tree that first boot wrote, so this is the ordinary path.
• Skill first on a fork that already runs: apply it, then stamp the group.

Either way the skill's last step restarts the agent containers as well as the service: a container that keeps running answers with the code it booted with. groups in mwe.json stays empty — it is an opt-out list, and what makes a group ask for the memory is this plugin.

The bridge README has the full walkthrough.

No mcp.json

The memory server is deliberately not declared as an MCP server here. Two reasons, both structural:

• The bridge speaks for one sender per turn (X-MWE-Act-As), and a static header in mcp.json cannot vary per sender.
• nanoclaw refuses a plugin MCP server whose URL host reaches the container host (src/templates/mcp.ts), which is exactly where a self-hosted memory server sits.

mwe_search and mwe_dashboard_link are registered by the skill, on the container's own in-process MCP server, with the speaker's act-as attached.