memories-for-llms
ossLoRA-as-memory experiment: per-user durable memory lives in a rank-16 LoRA adapter on a Qwen3 student rather than in the prompt window.
The design is dual-tier: a SQLite fact store with RAG handles online recall, while a periodic QLoRA consolidation job rewrites the adapter offline using teacher-driven generative replay against catastrophic forgetting.
Teacher (Qwen 3.6 35B) generates replay pairs from new facts. The student (Qwen 3.5 9B, Q4_K_M) runs locally with the LoRA adapter loaded. The adapter snapshot ships with a metadata.json listing the fact IDs that contributed to its gradient updates, so future passes can roll back problematic consolidations.
Status: M0 (SQLite + RAG) complete. M1 (consolidation loop) in progress.
The full thesis, novelty bet, and experiment plan live in the project wiki.