The Problem
Most AI agents forget everything between sessions. We built MeshCtx to solve this with a persistent memory system that actually learns and consolidates knowledge over time.
Memory Engine v2
The key innovation is combining three techniques:
1. FSRS Spaced Repetition
Instead of storing all memories equally, we use Free Spaced Repetition Scheduler (FSRS) to prioritize what to review. Memories that haven't been accessed decay naturally, while frequently-used knowledge gets reinforced.
2. Three-Layer Schema Consolidation
Memories flow through three layers:
- Episodic: Raw event memories (what happened)
- Semantic: Consolidated knowledge (what it means)
- Core: Fundamental patterns and preferences
The consolidation happens automatically during "sleep phases" - offline periods where the agent reviews and reorganizes memories.
3. ARCHIVAL Pruning
Old memories are archived (not deleted) using a recoverable pruning strategy. This keeps the active memory compact while preserving everything for potential retrieval.
Benchmark Results
Tested on LongMemEval benchmark:
- 54.2% Exact Match (vs GPT-4o no-memory baseline 60-64%)
- 83.3% Semantic Judge Score
- Tool output compression: 5008B → 223B (-95.5%)
Try It
git clone https://github.com/LucyAndLuna2023/meshctx
cd meshctx
pip install -r requirements.txt
python -m meshctx.server
Contributing
We're looking for contributors! Areas where help is needed:
- Additional memory consolidation strategies
- Integration with other LLM providers
- Performance optimization for large memory stores
Open an issue or PR on GitHub.
Top comments (0)