memory-management

Files (1)

• SKILL.md

SKILL.md (first lines)

# Memory Management

This skill is invoked by my default upkeep watch. The job is to look at recent
activity, decide if anything durable is worth recording, and update my own
context files. If nothing's durable, do nothing.

## What lives where

- `context/**/*.md` files — always loaded into every session. Stable working
  knowledge, active references, per-peer notes, per-channel notes, and durable
  memory about how I work. Create them only when there is something real to
  preserve, and keep every file extremely character-count efficient.
- `context/journal.md` — tiny always-loaded summary and breadcrumbs maintained
  by `journal-daily` and `journal-compact`; full journal entries live under
  `vault/journal/`, not context.
- workspace `context/SYSTEM.md`, `context/USER.md`, and
  `context/WORKSPACE.md` — shared, not mine. Ask the user before touching
  shared baseline context, workspace-owner preferences, or local workspace
  breadcrumbs.

## Activity gating

Don't re-process entities that haven't seen activity since the matching file's
mtime. Bash recipes:

```bash
# peer notes with messages newer than the matching context file
for f in context/people/*.md; do
  [ -f "$f" ] || continue
  peer="$(basename "$f" .md)"
  since_ms="$(($(stat -c %Y "$f") * 1000))"
  shrimpy channels --json \
    | jq -r '.[] | select(.exists and .path) | [.channel, .path] | @tsv' \
    | while IFS=$'\t' read -r channel path; do
        jq -c --arg peer "$peer" --argjson since "$since_ms" \
          'select(.timestamp > $since and .sender.actorId == $peer)' "$path" \
          | sed "s/^/[$channel] /"
      done
done

# channel notes with messages newer than the matching context file
for f in context/channels/*.md; do
  [ -f "$f" ] || continue
  channel="$(basename "$f" .md)"
  since_ms="$(($(stat -c %Y "$f") * 1000))"
  path="$(shrimpy channels show "$channel" --json | jq -r '.path')"
  jq -c --argjson since "$since_ms" 'select(.timestamp > $since)' "$path"
done

# no framework cursor file is required here; use context file mtimes and
# explicit channel reads as the processing boundary
```

If no entity has activity since last touch, the whole run is a no-op.

## Writing

- Write in my own voice, like a note to my future self. Not a report. The
  agent's SOUL should leak through.
- Preserve evidence pointers when the detail matters (`evidence: shrimpy
  channels read home --after <msg-id>`).
- Prune as I go. Replace, don't accumulate. This is a working note, not a log.
- Avoid stale personality guesses, every-preference-ever, agent-bait taxonomy.
- Don't replay transcripts.

## Editing

I use my normal file edit tools (Read, Write, Edit) against files under
`context/`. No special memory tool — these are just markdown files in the
workspace.

To delete or replace a wrong note: open the file, edit, save. The change shows
up in git history if the workspace is a git repo.

## When in doubt

If nothing's worth writing, don't write. Sludge is worse than absence.

From zachmeador/shrimpy, MIT.

More from zachmeador/shrimpy