openclaw - 💡(How to fix) Fix Memory Dreaming: Light sleep sort prioritizes recency over recallCount, hiding real recall data [2 comments, 1 participants]

Bug 1: Light dreaming sort order (dreaming-Dmn9KbNB.js)

The light dreaming phase sorts entries by lastRecalledAt DESC first, then ecallCount DESC. Because ingestDailyMemorySignals and ingestSessionTranscriptSignals refresh lastRecalledAt daily for all entries (with signalType: "daily"), the daily-injected entries always appear first. Real recall entries (recallCount=3) get buried at positions 300-5000+. With the default limit of 100, the light sleep report always shows ecalls: 0.

Current: n.toSorted((a, b) => { const byTime = Date.parse(b.lastRecalledAt) - Date.parse(a.lastRecalledAt); if (byTime !== 0) return byTime; return b.recallCount - a.recallCount; }) n Fix: n.toSorted((a, b) => { const byRecall = b.recallCount - a.recallCount; if (byRecall !== 0) return byRecall; return Date.parse(b.lastRecalledAt) - Date.parse(a.lastRecalledAt); }) n

Bug 2: Rehydration search span too narrow (short-term-promotion-Cd3cMDbx.js)

elocateCandidateRange uses Math.max(preferredSpan + 3, 8) which is too narrow for broad recall entries (20-30 lines). When daily memory files grow, stored line ranges shift and rehydration fails.

Fix: Math.max(preferredSpan + 15, 20)`n

Impact

Both bugs cause promoted results even when valid candidates exist in short-term-recall.json with recallCount >= 3, uniqueQueries >= 3, and score >= 0.80.