Memory has three tools that make sense once you have a week of data in it:

• nex_synthesize — clusters learnings of the same category (pattern, mistake, decision, etc.) and asks the model to draw a higher-level finding from the cluster. "5 mistake-learnings about Postgres pool sizing" → one synthesis: "Pattern: I underestimate connection counts in new services. Default to pool=10 from now on."
• nex_reflect with days: 7 — generates reflections over the past week. What patterns emerged? What decisions am I making implicitly that I should make explicit? What gaps in my knowledge graph showed up multiple times?
• nex_communities — runs an edge-weighted community detection on your knowledge graph. Surfaces clusters of entities + observations that belong together. Useful for "what topics am I actually working on?" — the answer often surprises you.

A weekly review skill bundles these three.

If your memory is more than a few months old, also run nex_decay stats. It reports which learnings have not been read in 45+ days. Two outcomes:

• nex_learn_archive for the ones that are obviously stale — old project decisions, abandoned approaches.
• nex_learn_update for the ones that are still true but never came up because the wording is different from how you search now — fix the wording so future searches find them.

Including decay scan in the weekly review is the difference between memory that grows usefully and memory that grows into a junk drawer.

If you want the canonical version:

Install the weekly-review skill from aiguide

aiguide_install_skill { slug: "weekly-review" } if available, or skip ahead to writing it yourself in Step 6.

Run aiguide_validate_step. The validator confirms ~/.claude/skills/ has at least one entry. After you install or write the skill, that passes.

Weekly review

Output (one screen, dense):

Weekly review — Apr 19-25
─ Sessions: 23 (12 productive, 4 deep-work, 7 firefighting)
─ Top topics by volume: Postgres pool sizing, OAuth provider apps, recipe MDX format
─ Decisions made: 4 (pricing tier consolidation, RLS deferred, GEO architecture B, debrief skill default)
─ Patterns synthesized:
  - Underestimating connection counts in new services → from now default pool=10
  - OAuth callback port collision (Codex 1455) → install Desktop first, VS Code second
─ Stale learnings: 6 archived (ChatGPT-3.5 prompts), 4 updated (better keywords)
─ Knowledge gaps: "Stripe webhook deduplication" — referenced 4x, no entity yet — create one
─ Open going into next week: 2 follow-ups, 1 blocked decision

That is the entire week in a structure you can absorb in 90 seconds. If something on it is interesting, you click into the underlying entries with a follow-up question.

---
description: Weekly review — synthesize the week into one screen. Patterns plus anomalies plus stale memory plus open items for next week. Use Friday afternoon or end of week.
---

When the user says "weekly review", "wrap the week", or "/weekly":

1. Call these in parallel:
   - nex_reflect days=7
   - nex_synthesize category=mistake
   - nex_synthesize category=pattern
   - nex_communities (latest week)
   - nex_decay stats

2. From the decay stats, identify candidates for archive (>45 days unused) vs update (still relevant, bad keywords). Ask the user which to do — do not auto-archive.

3. Roll into 8-12 lines max. Lead with session count + topic distribution. End with open items for the new week.

4. No emojis. No headers. Dense factual lines.

The "do not auto-archive" rule is important. Memory is yours, deletion should be confirmed. The weekly review surfaces decay candidates; you approve or skip. Two minutes of attention per week keeps memory clean.

Recipe 5.2 covered scheduling. To put weekly review on auto:

/schedule
- Name: weekly-review
- Schedule: 0 17 * * 5 (Friday 17:00)
- Prompt: Run the weekly-review skill. If patterns or stale learnings come up that need decisions, post the summary to Telegram so I see them on my phone.

The Telegram pings the patterns to your phone — you read them away from the terminal, decide whether to dig in Monday morning.

What you gain from doing this

The first time you run a weekly review you are surprised by something. Either by what dominated your week (you thought you were working on X but most of your sessions were Y), or by a pattern you had not noticed (the same root cause behind multiple bug fixes). That insight pays for the entire setup.

After a few weeks the surprises decrease — you have absorbed the patterns. The review becomes a quick pulse-check instead of a revelation. That is the goal: weekly reviews are how you keep your model of your own work accurate.

When NOT to bother

If you are early in adopting persistent memory (under 4 weeks), there is not enough data for synthesis to be interesting. Just keep daily debriefs going (Recipe 5.3). Add weekly reviews after the first month, when there is something to look back at.