Performance

Optimize Mnemosyne for your workload.

Benchmarks

All benchmarks run on a single core, SQLite with sqlite-vec:

Optimization Strategies

1. Batch Writes

# Bad: Individual writes (separate transactions)
for item in items:
  mem.remember(item)

# Good: Batch via remember_batch
mem.beam.remember_batch([
  {"content": item, "source": "batch", "importance": 0.5}
  for item in items
])

2. Enable WAL Mode

SQLite WAL mode improves concurrent read/write performance. Enable it manually on your database file:

sqlite3 ~/.hermes/mnemosyne/data/mnemosyne.db "PRAGMA journal_mode=WAL;"

3. Embedding Caching

Mnemosyne automatically caches embedding queries via LRU cache (512 entries). Repeated queries in agent loops are near-instant. No configuration needed.

4. Working Memory Limits

Keep Working Memory lean. The default maximum is 10,000 entries, controlled via environment variable:

export MNEMOSYNE_WM_MAX_ITEMS=5000  # Lower limit for faster queries
export MNEMOSYNE_WM_TTL_HOURS=12    # Shorter TTL for more aggressive eviction

5. Regular Consolidation

Run consolidation periodically to promote working memories to episodic and maintain query performance:

# Standalone CLI
mnemosyne sleep

# Or programmatically
mem.sleep()

Memory Usage