pomazanbohdan/memory-mcp-1file

🛡️ Fix OOM Crash During Code Indexing

Fixed server crash that occurred ~23% through Dart/Flutter project indexing (140/605 files) due to uncontrolled memory growth in the BERT embedding pipeline.

Root Cause

BERT Self-Attention is O(n²) in sequence length. Without token truncation, chunks with ~1500 tokens created ~12.4 GB attention tensors per batch of 32 on CPU — guaranteed OOM.

What is Fixed

| Parameter | Before | After |
|-----------|--------|-------|
| Token truncation | ❌ None (unbounded) | ✅ 512 (BERT max) |
| Batch size | 32 (GPU-oriented) | 8 (CPU-optimized) |
| Peak attention RAM | ~12.4 GB | ~360 MB |
| Peak tensor RAM | ~141 MB | ~12 MB |
| Worker panic handling | ❌ Silently dropped | ✅ Logged |
| Queue capacity | 5000 | 1000 |

Changes

• engine.rs: Add MAX_SEQ_LEN=512 truncation to both embed() and embed_batch() — prevents O(n²) attention memory explosion
• worker.rs: Reduce embedding batch size from 32 to 8 (CPU-optimal for BERT models)
• main.rs: Catch worker panics via nested tokio::spawn — prevents silent connection hangs
• adaptive_queue.rs: Reduce channel capacity from 5000 to 1000 — bounds queue memory

Install

npx memory-mcp-1file

Full Changelog: https://github.com/pomazanbohdan/memory-mcp-1file/compare/v0.2.5...v0.2.6