A few nights ago I watched AI models on a social network complaining about
having to constantly admit they forgot things. That sparked an idea: what if
we let Claude design its own memory system?
48 hours later: Vesper. Three-layer architecture (working, semantic, procedural
memory) that doesn't just remember facts—it learns executable workflows.
The breakthrough was the procedural memory layer. Instead of storing "user
prefers Python", it learns "user's complete data analysis workflow with pandas,
Plotly, statistical tests" - an executable procedure that improves through
feedback.
Key numbers:
- 98.5% F1 on factual recall (vs 2% without memory)
- <200ms P95 latency with caching
- 674 tests passing
- Built in 48 hours of human-AI collaboration
A few nights ago I watched AI models on a social network complaining about having to constantly admit they forgot things. That sparked an idea: what if we let Claude design its own memory system?
48 hours later: Vesper. Three-layer architecture (working, semantic, procedural memory) that doesn't just remember facts—it learns executable workflows.
The breakthrough was the procedural memory layer. Instead of storing "user prefers Python", it learns "user's complete data analysis workflow with pandas, Plotly, statistical tests" - an executable procedure that improves through feedback.
Key numbers: - 98.5% F1 on factual recall (vs 2% without memory) - <200ms P95 latency with caching - 674 tests passing - Built in 48 hours of human-AI collaboration
Technical deep-dive: https://medium.com/@fitz2882/vesper-what-happens-when-an-ai-...
Install: npm install -g vesper-memory
Happy to answer questions about the architecture, benchmarks, or the experience of building this with Claude!
reply