Persona — Learnable LLM Memory Layer
Writeup in progress. Full research notes covering the salience-vs-frequency problem, contrastive training signal design, Qdrant integration, and benchmark methodology coming soon.
All projectsProject
Persona — Learnable LLM Memory Layer
Research-stage learnable scoring layer that separates access frequency from semantic salience in LLM memory systems — the failure mode that causes drift and recall failures in Mem0, Letta, and Zep on long-horizon tasks.
llmmemoryagentsopen-source
Problem
Existing memory systems (Mem0, Letta, Zep) conflate access frequency with semantic salience, causing memory drift and ranking failures. High-frequency but low-salience memories crowd out critical context, degrading agent performance over long-horizon tasks.
Solution
Ingestion of ChatGPT and Claude exports, episode segmentation, SQLite + sqlite-vss with bge-small-en-v1.5 embeddings, and a FastAPI intercepting proxy. On top of that, a specified write-policy scorer, mode classifier and retrieval reranker — under 10M parameters total — that re-rank memories by predicted future utility rather than historical access count.
Impact
Research-stage, and deliberately pre-registered: a locked 55-probe evaluation set (30 identity, 10 transactional, 15 refusal) and three baselines exist before any model was trained. No Persona-versus-baseline result is claimed yet — the three learnable heads are designed and specified, not trained. Week 3.5 of a 12-week proof of concept.
Stack
PythonPyTorchQdrantpydanticasyncio