Long-term memory quality in LLM systems is often limited by the model that executes the write path: generic LLMs either over-store, miss salient facts, or fail to reconcile conflicting updates. This project targets that bottleneck by training a dedicated memory-writer that learns framework-specific protocols and optimizes downstream answerability across multiple deployed memory frameworks.
Key Capabilities
- Trained for memory writing, not general chat: a two-stage pipeline (Memory On-Policy Distillation followed by Attributed Memory Reinforcement Learning) teaches framework-specific write behaviors and then optimizes chunk-level credit assignment and bank-level utility.
- Cross-framework transfer: validated inside four memory systems (Mem0, MemOS, Hy‑Memory, held-out EverMemOS) and ranks top on a fixed 20% benchmark matrix with a 55.49% macro average versus 54.62% for a strong baseline (Claude Opus 5), indicating robust out-of-distribution transfer.
- Efficiency and deployment: implemented as a compact ~3B active-parameter writer (30B A3B MoE family) that replaces heavier general-purpose LLMs inside memory frameworks and is compatible with vLLM-based serving and OpenAI-compatible APIs.
- Fine-grained memory controls: training includes dense per-chunk probes, attributed cross-chunk rewards, and over-storage constraints to reduce redundant or low-utility memory writes.
Who it's for and tradeoffs
Great fit if you build or research retrieval-augmented agents and need a specialized component that extracts, summarizes, and reconciles persistent user or task state across long horizons; especially useful when you want predictable, bank-level improvements in answerability without running a full general-purpose LLM on every write. Look elsewhere if you need a general conversational model, multimodal capabilities, or an off-the-shelf reader/answering model—this is explicitly a write-path specialist and assumes a separate answer model and retrieval stack. Training and evaluation focus on benchmarked memory utility rather than open-ended generation quality, and the model is designed to be integrated into Hy‑Memory/Mem0/MemOS-style frameworks rather than used standalone as a chat agent.
