Consolidate Memory

Why this exists, and what I'd change

Why it exists. Single-file rolling memory gets noisy fast. Long-term decisions get buried under last week's standup chatter. This skill splits memory into three files (recent, long-term, project) and promotes the durable stuff up the chain.

Design tradeoffs.

• Three memory files instead of one. Recent (48h rolling), long-term (distilled facts and preferences), project (active initiative state). Cost: more files to grep, and promotion logic is heuristic and occasionally wrong.
• Nightly cron promotion via launchd, not in-session. Promotion runs at 2am local time. Cost: same-day insights don't move into long-term until tomorrow. If the cron breaks, memory drifts silently.
• Reads from ~/.claude logs, not a dedicated event stream. I chose to scrape what's already there. Cost: depends on Claude Code's log format staying stable. Will break if Anthropic changes it.

What I'd change. Add an explicit "remember this" command that promotes a fact immediately to long-term, instead of waiting for the cron to maybe catch it.

Consolidate Memory

Keep a persistent memory layer up to date from real conversation history.

Scope

This skill does four things:

1. Reads local conversation logs from ~/.claude (last 24h for extraction).
2. Updates rolling context in memory/recent-memory.md (48h window).
3. Updates memory/project-memory.md (active initiatives, open decisions, next actions).
4. Promotes durable facts/preferences/patterns into memory/long-term-memory.md.

Commands

/consolidate-memory run
/consolidate-memory status
/consolidate-memory install-nightly

Equivalent CLI:

python scripts/memory/consolidate_memory.py --run

Files

• memory/recent-memory.md - Rolling 48-hour memory
• memory/long-term-memory.md - Distilled durable memory
• memory/project-memory.md - Active project state

Legacy mirrors are kept in:

• long-term-memory.md
• context-library/pm-os-learning-log.md

Promotion Rules

Promote from recent -> long-term when signals look durable:

• repeated mentions (frequency threshold)
• explicit preference language (prefer, always, never, tone, style)
• durable operational facts (scheduled, deadline, recurring workflows)

Nightly Automation (macOS)

Install/update launchd job:

bash scripts/memory/install_consolidate_memory_launchd.sh

This writes:

• ~/Library/LaunchAgents/com.pmos.consolidate-memory.plist

And schedules a nightly run at 2:20 AM local time.

Output Expectations

After each run, script prints a JSON summary:

• number of source files read
• extracted decisions/preferences/facts
• promoted facts/preferences/patterns