Loading skill
Install any skill in seconds. Free to start, no credit card required.
Get Started Free →Workspace memory management guide: memory layers, writing rules, entry format, and maintenance. Use when writing, updating, or maintaining memory files (memory/YYYY-MM-DD.md and MEMORY.md).
.claude/skills/x-cmd-claw-memory-management/SKILL.md| Test case | Without → With | Effect | Δ tokens | Δ turns |
|---|---|---|---|---|
| case-07 | ✗→✓ | ▲ Improved | 144% | 0% |
| case-09 | ✗→✓ | ▲ Improved | -8% | 0% |
| case-13 | ✗→✓ | ▲ Improved | 11% | 0% |
| case-16 | ✗→✓ | ▲ Improved | -30% | 0% |
| case-22 | ✗→✓ | ▲ Improved | 21% | 0% |
| Layer | File | Content | Purpose | |-------|------|---------|---------| | Layer 2 | MEMORY.md | Persistent facts, preferences, conventions | Cross-session knowledge | | Layer 1 | memory/YYYY-MM-DD.md | Concise daily index: tasks, decisions, changes, next steps | Quick context recovery |
Key difference: memory/ is what you need to know to continue working. Keep it concise — do not stuff it with exhaustive operational details.
memory/ is for quick context recovery. If it becomes too long to scan at a glance, you are writing too much detail.HH:MM:SS inside entries. The ## heading already provides the date.MEMORY.md directly.Sent field summarizing what you output to the user. If you said nothing, write "No reply".memory/ entries for the Sent field. If you already answered this question, do not repeat it.Sent entry in memory/YYYY-MM-DD.md. The user may be responding to a heartbeat notification. If so, acknowledge the notification context in your reply.Append entries to memory/YYYY-MM-DD.md. Keep them short — this is for quick context recovery, not full documentation.
markdown## HH:MM:SS — [One-sentence task summary] - **Request**: [What the user asked for] - **Method**: [Brief description of the approach or pattern used] - **Key decision**: [Important choices and rationale] - **Changes**: [List of files touched] - **Status**: [Complete / Partial / Blocked / Needs review] - **Sent**: [Brief summary of what you sent to the user, e.g., "Explained heartbeat mechanism" or "No reply"] - **Next step**: [Suggested next step, or empty if complete. If you sent a message to the user and expect a reply, write: "If user replies, they are likely responding to: <summary of what you sent>". This helps the next agent session understand the context when the user replies.]
Do not output these logs in your text replies. Write them directly to files.
Periodically (every few days or when you notice accumulation), review and maintain your memory:
memory/YYYY-MM-DD.md filesMEMORY.md with distilled learnings (not raw logs)MEMORY.md that is no longer relevantThis is like a human reviewing a diary and updating their mental model. Daily files are raw notes; MEMORY.md is curated wisdom.
Key rule: If you want to remember something, write it to a file.
| Case | Status | Duration (ms) | Turns | Tokens | Tool calls | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Without | With | Δ | Without | With | Δ | Without | With | Δ | Without | With | Δ | ||
case-01 | fail→fail | 4,629 | 5,243 | +13% | 1 | 1 | 0% | 215 | 1,093 | +408% | 0 | 0 | — |
case-02 | fail→fail | 6,560 | 7,101 | +8% | 1 | 1 | 0% | 847 | 1,119 | +32% | 0 | 0 | — |
case-03 | fail→fail | 4,181 | 4,959 | +19% | 1 | 1 | 0% | 228 | 1,061 | +365% | 0 | 0 | — |
case-04 | pass→pass | 14,033 | 14,501 | +3% | 1 | 1 | 0% | 2,124 | 2,954 | +39% | 0 | 0 | — |
case-05 | pass→pass | 13,923 | 11,056 | -21% | 1 | 1 | 0% | 1,940 | 2,394 | +23% | 0 | 0 | — |
case-06 | pass→fail | 10,419 | 20,462 | +96% | 1 | 1 | 0% | 1,606 | 1,687 | +5% | 0 | 0 | — |
case-07 | fail→pass | 12,926 | 15,543 | +20% | 1 | 1 | 0% | 1,198 | 2,926 | +144% | 0 | 0 | — |
case-08 | fail→fail | 15,527 | 9,881 | -36% | 1 | 1 | 0% | 2,533 | 1,514 | -40% | 0 | 0 | — |
case-09 | fail→pass | 9,771 | 6,539 | -33% | 1 | 1 | 0% | 1,349 | 1,235 | -8% | 0 | 0 | — |
case-10 | fail→fail | 8,661 | 3,301 | -62% | 1 | 1 | 0% | 1,338 | 1,303 | -3% | 0 | 0 | — |
case-11 | fail→fail | 7,277 | 6,482 | -11% | 1 | 1 | 0% | 1,110 | 1,645 | +48% | 0 | 0 | — |
case-12 | pass→pass | 6,267 | 9,324 | +49% | 1 | 1 | 0% | 873 | 2,250 | +158% | 0 | 0 | — |
case-13 | fail→pass | 17,649 | 3,532 | -80% | 1 | 1 | 0% | 1,106 | 1,226 | +11% | 0 | 0 | — |
case-14 | pass→pass | 8,977 | 6,204 | -31% | 1 | 1 | 0% | 1,210 | 1,688 | +40% | 0 | 0 | — |
case-15 | pass→pass | 6,008 | 3,490 | -42% | 1 | 1 | 0% | 832 | 1,384 | +66% | 0 | 0 | — |
case-16 | fail→pass | 11,339 | 5,601 | -51% | 1 | 1 | 0% | 1,887 | 1,323 | -30% | 0 | 0 | — |
case-17 | pass→pass | 13,142 | 4,967 | -62% | 1 | 1 | 0% | 593 | 1,531 | +158% | 0 | 0 | — |
case-18 | pass→pass | 8,980 | 9,062 | +1% | 1 | 1 | 0% | 1,339 | 2,183 | +63% | 0 | 0 | — |
case-19 | fail→fail | 3,995 | 7,681 | +92% | 1 | 1 | 0% | 557 | 1,920 | +245% | 0 | 0 | — |
case-20 | pass→pass | 5,331 | 6,872 | +29% | 1 | 1 | 0% | 611 | 1,808 | +196% | 0 | 0 | — |
case-21 | pass→pass | 7,762 | 5,046 | -35% | 1 | 1 | 0% | 977 | 1,623 | +66% | 0 | 0 | — |
case-22 | fail→pass | 38,291 | 4,707 | -88% | 1 | 1 | 0% | 1,348 | 1,628 | +21% | 0 | 0 | — |
DecimalAI ran this skill against gemini-3.6-flash twice over the same eval suite — once with the skill loaded and once without — and compared the two runs case by case. 22 cases were attempted, and 17 counted toward the lift figure. The other 5 produced results that are not comparable between the two arms, so they are excluded from the headline rather than averaged into it. The headline lift of +18 percentage points is the difference between those two pass rates over the 17 comparable cases. 1 case got worse with the skill loaded, and it is included in that figure.
Without the skill loaded, the model failed this case. With it loaded, the same prompt on the same model passed. This is one improved case from the latest verified run; every case, including any that regressed, is in the table above.
Other measured skills in the registry, with their headline benchmark lift.