memory-distill

Memory distill

The log-correction-signal hook captures every prompt where Marco corrected behaviour or stated a preference, at zero token cost. This skill turns that raw log into memory files. Run it periodically, not every session.

Why it works this way

Auto-extracting memories at the end of every session burns tokens and produces low-grade noise ("user wanted the button blue"). Corrections are the high-signal moments: they mark where the assistant's default diverged from what Marco actually wants. Logging is free; distilling is deliberate.

Workflow

Step 1 — Read the log

wc -l ~/.claude/memory-signal.jsonl 2>/dev/null || echo "no signal log yet"
cat ~/.claude/memory-signal.jsonl

Each line is { at, cwd, session, prompt, hits }. If the file is missing or empty, say so and stop; don't invent memories.

Step 2 — Read what's already stored

Memory lives per-project at ~/.claude/projects/<slug>/memory/. Read the relevant MEMORY.md index and any file that might already cover a theme. You are updating, not appending duplicates.

Step 3 — Group into themes

Cluster the raw prompts. A theme needs at least two independent occurrences, or one very explicit standing instruction ("from now on…"). Discard:

• One-off task corrections with no general rule ("no, the other file")
• Anything already covered by CLAUDE.md, a skill, or the repo's own docs
• Preferences derivable from the code itself (the marco-code-style skill covers style; don't duplicate it here)

For each surviving theme, write down the why. A memory without a reason gets misapplied later.

Step 4 — Propose, don't write

Present each candidate as:

### <slug>
**Type:** feedback | project | user | reference
**Claim:** <one line>
**Why:** <the reason it matters>
**Evidence:** <2-3 quoted fragments from the log, with dates>
**Action:** new file | update <existing-file.md> | supersedes <old-file.md>

Then ask which to keep, using a numbered list so a numbered reply maps by index.