Bug Report

Description

openclaw memory index --force consistently fails with Memory index failed (main): fetch failed on a workspace with 1400+ .md files. After 2 failures, batch mode is permanently disabled (Batch: disabled (failures 0/2)), making re-indexing impossible without restarting gateway.

Environment

• OpenClaw version: 2026.3.24
• OS: macOS 26.3 (arm64)
• Node: v25.8.1
• Memory plugin: memory-core (stock, loaded)
• Embedding provider: openai / text-embedding-3-small
• OpenAI tier: Tier 3

Steps to Reproduce

1. Have a workspace with 1400+ .md files in memory/
2. Run openclaw memory index --force --verbose
3. Observe 100-200+ [memory] embeddings: batch start lines
4. Process fails with Memory index failed (main): fetch failed

Expected Behavior

Indexing completes successfully with proper retry/backoff on rate limits.

Actual Behavior

• Indexer fires all batches in parallel (no concurrency limit)
• One batch fails (rate limit or timeout)
• Entire transaction rolls back
• Batch: disabled (failures 0/2) — batch mode disabled permanently until gateway restart
• Subsequent runs fail immediately

Logs

[memory] sync: indexing memory files
[memory] embeddings: batch start   (x200+)
[memory] embeddings rate limited; retrying in 574ms
Memory index failed (main): fetch failed

[memory] sync failed (session-start): TypeError: fetch failed
[memory] sync failed (search): TypeError: fetch failed

Workarounds Tried

• Reducing workspace size (moved sessions/, backups/, _nonmd/ out of memory/) — no effect
• Clearing sqlite store — no effect
• Waiting for rate limits to reset — no effect
• Increasing OpenAI tier — no effect

Suggested Fix

1. Add concurrency limit to embedding batch processing (e.g. max 5 parallel batches)
2. Proper exponential backoff on 429/rate-limit responses
3. Do NOT disable batch mode globally after N failures — retry per-batch instead
4. Persist partial index progress so failed runs can resume