memory-flush

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:

1. Scan recent logs for entries worth promoting to long-term memory:

   • New lessons learned (errors encountered, workarounds found)
   • Topics covered (articles, digests) — add to the recent articles/digests tables
   • Features built or tools created
   • Important findings from monitors (on-chain, GitHub, papers)
   • Ideas captured that are still relevant
   • Goals completed or progress milestones

2. Check each candidate against existing MEMORY.md content — skip if already recorded.

3. Update memory:

   • Add brief entries to MEMORY.md
   • If a topic needs more detail, write to memory/topics/<topic>.md instead
   • Update the Recent Digests table with any new token-report or push-recap entries from logs

4. Rotate old entries to keep MEMORY.md under ~50 lines:

   • Skills Built table: keep the 10 most recent rows — remove older rows from the top
   • Recent Articles table: keep the 8 most recent rows — remove older rows from the top
   • Recent Digests table: keep the 6 most recent rows — remove older rows from the top

5. 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):

   • Skills Built rows: ≤ 280 chars per row body
   • Recent Articles rows: ≤ 220 chars per row body
   • Recent Digests rows: ≤ 180 chars per row body

   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.

6. Update the "Last consolidated" date at the top of MEMORY.md to today (${today}). Do this even if nothing else was promoted.

7. 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.)

8. Log what you promoted (or "MEMORY_FLUSH_OK: no new entries, date updated") to memory/logs/${today}.md.