rlm

name

rlm

description

Run the brain-agent RLM synthesizer on a task — an agent with a persistent Python shell, a CONTEXT state dict of chunk slots, and the full brain SDK. It routes across your registered brains (route/reward bandit), builds new specialist brains on the fly, and returns FINAL. Use for any task over context too large or too structured for one window — analysis over big corpora, multi-lens reviews, judged sweeps.

RLM — the brain-manager agent as a callable script

One invocation = one task handed to the synthesizer agent. It manages its own work: engineers CONTEXT slots, tiles/lenses the material, routes to existing registered brains or designs new ones, and sets FINAL.

python -m brain_agent.orchestrator "TASK" \
  [--kernel NAME]            # persistent runtime context (default: rlm)
  [--reset]                  # fresh kernel + CONTEXT
  [--bind slot=/path/file]   # load a file into a CONTEXT slot (repeatable)

Requires: MINIMAX_API_KEY (or heaven-routed provider), HEAVEN_DATA_DIR (brain registry + bandit ledger), optionally BRAIN_CONTEXT_DIR, BRAIN_TRACE_DIR (call-graph), BRAIN_KERNEL_DIR.

The kernel persists between invocations: calling again with the same --kernel resumes with all slots, vars, and brains intact. The bandit ledger (route/reward) persists across ALL kernels — routing improves globally.