Ruflo 3.41.2 stops a memory curate from shrinking your index

What Ruflo is

If you already use Claude Code, Ruflo is the layer underneath it that keeps work organised. Agents self organise into swarms, tasks get routed, and what worked last week is still there next week because memory persists across sessions.

You do not have to learn the tool surface to benefit from it. After the one time init, you keep writing code the way you already do and the hooks handle routing, learning and coordination in the background.

What changed in 3.41.2

This is a patch release built from the issue backlog. The headline fix is issue #3224: the automatic memory bridge rebuilt MEMORY.md from its own topic files and overwrote whatever a human had written there. One reporter watched a curated index of 75 lines and 49 links collapse into a six line stub.

A curate may now grow or reorder the index. It may not shrink it. The release also covers daemon consent and sibling disclosure fixes, each with a regression test that fails without the change.

This is a fix to existing behaviour, not a new feature. If you never used the memory curator, nothing in your setup changes.

Get started

Node.js and a working Claude Code install. One command sets everything up:

npx ruflo@latest init wizard

Expected result: a scaffolded project with agents, commands, skills and hooks in place, and Ruflo registered so Claude Code can reach it.

To register Ruflo as an MCP server explicitly:

claude mcp add claude-flow -- npx ruflo@latest mcp start

There is also a Claude Code plugin track if you only want slash commands and agents:

/plugin marketplace add ruvnet/ruflo
/plugin install ruflo-core@ruflo

The plugin track and the CLI track expose different tool names. Plugin tools are namespaced as mcp__plugin_ruflo-core_ruflo__*; the CLI track uses the bare names. Pick one track per project.

Use it today

Practical case: a small team that keeps re-explaining the same architecture decisions to their coding agent. Run init in the repository, work normally for a week, and let the memory index accumulate the decisions that actually got made.

Acceptance test you can run: after init, write a few lines into MEMORY.md by hand, let a session run, then reopen the file. Your lines should still be there. Before 3.41.2 they could disappear.

Push it further

Experimental commentary. The interesting direction is federation: agents on separate machines coordinating without pooling their data. Ruflo ships a federation path and the x.ruv.io gateway, but treat cross machine coordination as something to test rather than assume.

Falsifiable test: run two federated instances, have one claim a resource with a short lease, and confirm the second cannot acquire it until the lease expires. If it can, your federation config is not doing what you think.