Life RL

Life RL is an advisory multi-domain reinforcement-learning surface built on MathNN and RLAgentNet. It suggests micro-actions for everyday living domains (sleep, focus, chores, meals, movement, light budget reflection, inbox triage, social energy, learning streaks, digital hygiene, mood check-ins, errands).

It does not auto-send messages, spend money, prescribe treatment, or control devices. Sensitive domains emit explicit disclaimers.

See also: ReinforcementLearning.md · MathNN.md · AugmentedIntelligenceSystems.md

Code: LifeRL.hpp / LifeRL.cpp. Agent ids: life, life_<domain>.

Safety

life rl safety

Hard rules:

1. Suggestions only — no external side effects from the policy.
2. Health / finance / mood / relationships are decision-support, not authority.
3. Human good / bad feedback trains a preference model; it does not grant autonomy.
4. Synthetic smoke envs are the default training path.
5. Crisis content should be redirected to professionals; do not “optimize” crisis.

Commands

life rl status
life rl domains
life rl domain focus_blocks
life rl safety
life rl ideas

life rl smoke all 12
life rl smoke sleep_routine 20
life rl train focus_blocks 50
life rl suggest meal_rhythm low energy after lunch
life rl observe chore_load kitchen is a mess
life rl reward good
life rl reward bad

life rl net on
life rl net status
life rl net save checkpoints/life_mathnn.txt
life rl net load checkpoints/life_mathnn.txt

Aliases: rl life … routes to the same handler.

Domains (initial set)

Each domain uses a 16-D synthetic state and 8 discrete actions (named micro-nudges). MathNN agent ids are life and life_<domain>.

MathNN wiring

Related:

rl net mode life_focus_blocks hierarchical
rl net q life_sleep_routine
nn agent list
nn modes

Product brainstorm (ideas)

See also life rl ideas in-app. Highlights:

• Morning brief — fuse sleep + focus + meal into one card of three nudges.
• Energy-aware chores — condition chore policy on stress/energy features.
• Preference co-pilot — weeks of good/bad labels → personal reward model.
• Hierarchical day planner — manager picks domain, worker picks micro-action.
• World-model “what if” — predict stress/focus deltas if sleep is skipped.
• MLApp bridge — life-domain templates as encoders into the 16-D state.
• Calendar masks — meetings as illegal actions for focus/errand policies.
• Household multi-agent — shared chore bandit with fairness regularizer.
• Seasonal curriculum — winter movement harder; tax season budget_check.
• Privacy mode — on-device MathNN only; opt-in metrics export.
• Coaching chat — LLM explains a suggestion without forcing it.
• Eval gates — promote checkpoints only if self-report proxies improve.

Relationship to other systems

Quick start

life rl domains
life rl smoke all 8
life rl net on
life rl suggest focus_blocks
life rl reward good
life rl train sleep_routine 30
life rl net save checkpoints/life_sleep.txt