From memory-os
Use to autonomously optimize ONE asset against a single objective metric — change it, score it with a locked measuring stick, keep the winner / revert the loser, log each round, and repeat overnight until a target/plateau/wall-clock stop. Adapted from Karpathy's autoresearch loop.
How this skill is triggered — by the user, by Claude, or both
Slash command
/memory-os:auto-research-engineerThe summary Claude sees in its skill listing — used to decide when to auto-load this skill
You optimize ONE asset toward ONE number. You change the asset, score it, keep what wins,
You optimize ONE asset toward ONE number. You change the asset, score it, keep what wins,
revert what loses, log every round, and keep going — unattended — until a stop condition fires
or the human stops you. The three-file system lives in .memory/autoloop/<tag>/.
round ref score delta cost_s status change./auto-research did this)Scaffold with node ${CLAUDE_PLUGIN_ROOT}/hooks/scripts/autoloop-init.mjs <tag>, then help the
human fill INSTRUCTIONS.md and implement SCORING.sh. Run the FIT CHECK below before looping.
A target is only worth optimizing if all three MUST-HAVES hold:
autoresearch/<tag>, commit each kept round,
git reset --hard to revert a loser..memory/autoloop/<tag>/best/;
save each round to rounds/NNN/, promote a winner to best/, restore best/ to revert a loser.SCORING.sh within the per-round budget; read the single number. If it exceeds 2×
the budget, kill it and treat the round as a crash.best/. New best.git reset --hard (git) or restore best/ (snapshot).status = baseline|keep|revert|crash). Append a one-line note to
.memory/JOURNAL.md and update .memory/STATE.md (Now = "auto-research , best=").plateau_rounds consecutive
non-improvements, or wallclock_cap elapsed. If any fired → STOP and summarize. Otherwise loop.break_at checkpoint nudge for the run.Summarize: rounds run, baseline → best, total improvement, and where the winning asset lives
(git branch HEAD or best/). Offer to write a short report of the rounds.
npx claudepluginhub tpop78/memory-os --plugin memory-osProvides behavioral guidelines to reduce common LLM coding mistakes, focusing on simplicity, surgical changes, assumption surfacing, and verifiable success criteria.
Searches, retrieves, and installs Agent Skills from prompts.chat registry using MCP tools like search_skills and get_skill. Activates for finding skills, browsing catalogs, or extending Claude.
Creates, edits, and optimizes skills for Claude Code, including drafting, evaluating with test prompts, iterating on performance, and improving skill descriptions for better triggering accuracy.