Every AI agent memory system I've used (Mem0, Honcho, Hindsight) has the same problem: they accumulate forever. Old facts pollute retrieval. More tokens → worse results. Your agent gets slower and dumber over time.

So I built recall-sqlite — a memory system that actually forgets.

The core idea: tiered storage. Memories are automatically promoted or demoted based on how often they're accessed.

• Hot tier (~500): ANN + keywords + FTS5 — fast full retrieval
• Warm tier (~5K): Keywords + FTS5 only — 66-99% less compute
• Cold tier (unlimited): Zero compute, auto-promoted when relevant

Key design decisions:

1. Zero LLM at query time (embedding model only, 150MB local)
2. No vector database (just SQLite + sqlite-vec)
3. Graceful degradation (keyword+FTS5 fallback when offline)
4. Automatic schema migration (no manual steps)
5. Single pip install, no API keys, no Docker

After 6 months of daily use with 1469 memories, latency stays at ~80ms and memory is fixed at ~1.5MB.

Just got PR'd into the Hermes Agent ecosystem.

pip install recall-sqlite
github.com/Jnocode/recall-memory