Summary

Fixes #12590

Problem

memoryFlush only fires on every other auto-compaction cycle. After a flush where compaction also completes, incrementCompactionCount bumps compactionCount from N to N+1, and the code assigns this new value to memoryFlushCompactionCount — synchronizing both counters. On the next cycle, shouldRunMemoryFlush sees them equal and skips the flush.

Fix

Stop overwriting memoryFlushCompactionCount with the post-increment value. The flush records the pre-increment compactionCount, so the next compaction cycle sees compactionCount > memoryFlushCompactionCount and correctly triggers the flush.

Change: 3 lines removed (the nextCount reassignment), let to const.

Reproduction and Verification

Before fix (main branch) — Bug reproduced:

PASS Cycle 1: first flush (compactionCount=1, no prior flush): got=true
FAIL Cycle 2: after flush+compaction (compactionCount=2, memoryFlushCompactionCount=2): got=false, expected=true
PASS Cycle 3: next compaction (compactionCount=3, memoryFlushCompactionCount=2): got=true
FAIL Cycle 4: after flush+compaction again (compactionCount=4, memoryFlushCompactionCount=4): got=false, expected=true

BUG CONFIRMED: memoryFlush skips cycles where counters are synchronized

After fix — All verified:

PASS Cycle 1: first flush triggers
PASS Cycle 2: flush triggers (compactionCount=2, flushCount=1)
PASS Cycle 3: flush triggers (compactionCount=3, flushCount=2)
PASS Cycle 4: flush triggers (compactionCount=4, flushCount=3)
PASS Dedup: same compactionCount, already flushed
PASS Under threshold: low tokens
PASS No entry: undefined

All checks passed

Effect on User Experience

Before fix: Users with memoryFlush enabled see memories saved only on ~50% of compaction cycles. Important context is silently lost.

After fix: memoryFlush fires reliably on every compaction cycle.

Testing

• 11 unit tests pass (memory-flush.test.ts)
• 8 integration tests pass (agent-runner.memory-flush.*.test.ts)
• Lint passes

Greptile Overview

Greptile Summary

This change fixes a regression where memoryFlush was only firing every other auto-compaction cycle. In src/auto-reply/reply/agent-runner-memory.ts, the flush now records memoryFlushCompactionCount using the pre-increment compactionCount and no longer overwrites it with the post-increment value returned by incrementCompactionCount, which previously synchronized the counters and caused the next cycle to be skipped.

Tests were updated to reflect the intended behavior by asserting memoryFlushCompactionCount remains at the pre-increment value even when compactionCount is incremented after a flush, and a new unit test was added to confirm shouldRunMemoryFlush triggers across successive compaction counts.

Confidence Score: 5/5

• This PR is safe to merge with minimal risk.
• The change is small and directly addresses the described counter-sync bug; it aligns with how incrementCompactionCount persists only the compaction increment while memoryFlushCompactionCount is persisted separately. Updated tests lock in the intended pre-increment recording behavior and cover the regression scenario at the unit and integration level.
• No files require special attention