vbario / sleeping-llm

What is sleeping-llm?

Sleeping-llm is a Python-based local language model AI that builds persistent memories from conversations by editing weights during "wake" chats and consolidating them via "sleep" cycles. Users chat normally, facts get injected instantly for recall, then sleep commands like /sleep or /nap audit, refresh, and transfer knowledge to stable long-term storage—no databases or retrieval needed. It handles up to 60 facts on 70B models with full recall post-sleep, running on MacBooks or H100s.

Why is it gaining traction?

Unlike RAG systems that stuff context or fetch from external stores, this embeds knowledge directly in weights for zero-latency recall and true generalization across conversations. Developers notice the neuroscience-inspired rhythm: fast wake edits degrade over time, sleep restores without PPL spikes, enabling unbounded capacity as short-term edits dissolve into fused long-term ones. Early benchmarks show 100% chat recall after 2-3 cycles, even as github language model unavailable issues plague Copilot alternatives.

Who should use this?

AI hobbyists running local Llama models who want conversation editing without losing facts between sessions. Researchers testing continual learning on edge hardware like M3 Macs, especially for language model ai experiments on memory retention. Indie devs building personalized copilots that evolve from user interactions, bypassing github language statistics showing Python's dominance in such tools.

Verdict

Worth forking for local LLM tinkerers—solid docs, 5 papers, and repro scripts despite 12 stars and 1.0% credibility score signaling early maturity. Test on 3B for quick wins, but expect tweaks for production as alignment taxes hit larger models.