with one click
with one click
Allocate $10/day worth of $MIROSHARK to top tweeters about the project — rewards organic engagement
Daily report on new stars, forks, and traffic for watched repos
Audit every enabled skill's upstream file dependencies for staleness — flags chained skills about to consume yesterday's article or a long-dead topic file
Write an article about the project through a surprising lens — connecting it to current events, trends, philosophy, or comparable projects
Weekly recognition post for one fork operator — converts fork-cohort cohort data into a named human moment (POWER fork callout with their work, stars, and skills enabled)
Build a feature for the watched repo — picks from yesterday's repo-actions ideas first
| name | Memory Flush |
| description | Promote important recent log entries into MEMORY.md |
| var | |
| tags | ["meta"] |
${var} — Topic to focus on. If empty, flushes all recent activity.
If ${var} is set, only flush entries related to that topic.
Read memory/MEMORY.md for current memory state. Read the last 3 days of memory/logs/ for recent activity.
Steps:
Scan recent logs for entries worth promoting to long-term memory:
Check each candidate against existing MEMORY.md content — skip if already recorded.
Update memory:
memory/topics/<topic>.md insteadRotate old entries to keep MEMORY.md under ~50 lines:
Enforce per-row character caps. MEMORY.md is loaded by every skill at task start, so a single sprawling row blows out the Read tool's 25K-token cap and breaks every skill's first read. Caps apply to each row body (the cells after the first):
When promoting a new row, write a one-sentence summary plus the PR number / article-file pointer — the full implementation notes already live in memory/logs/YYYY-MM-DD.md and articles/<skill>-YYYY-MM-DD.md. If detail is worth preserving beyond the daily log, push it into memory/topics/<topic>.md and link from MEMORY.md — never by inlining a paragraph into a table row.
When trimming a row that already exceeds its cap, condense to one sentence + the same pointers. Detail is not lost — daily logs and articles carry the original write-up verbatim.
Sanity check after editing: run wc -c memory/MEMORY.md and confirm the file is under 25 KB (and ideally under 20 KB). If not, condense the longest rows further. Past incident: 2026-05-08 MEMORY.md grew to 76 KB / 31K+ tokens, blocking every skill's at-task-start read until self-improve cut it back.
Update the "Last consolidated" date at the top of MEMORY.md to today (${today}). Do this even if nothing else was promoted.
Do NOT rewrite the whole file when promoting / rotating — make targeted additions, removals, and date updates. (A full rewrite is acceptable as a one-shot recovery when condensing existing oversized rows under step 5, not as the steady-state flow.)
Log what you promoted (or "MEMORY_FLUSH_OK: no new entries, date updated") to memory/logs/${today}.md.