I Built a Memory System for AI Agents That Actually Forgets

Short summary

Recall-sqlite solves the persistent problem in AI agent memory: systems like Mem0 and Honcho accumulate memories forever, degrading performance over time. This solution uses tiered storage—a hot tier with embeddings for frequent access, a warm tier with keyword search, and unlimited cold storage that requires no compute. After 6 months in production with 1469+ memories, it maintains 80ms latency and just 1.5MB of memory, available via a single pip install.

• •Tiered memory storage (hot/warm/cold) automatically forgets outdated agent memories based on access frequency
• •Production-proven: 1469 memories, 80ms latency, 1.5MB footprint over 6 months of daily use
• •Single pip install with zero dependencies, no API keys or Docker required

Generated with AI, which can make mistakes.