How Mind Lab Trained a Trillion‑Parameter Agentic Memory with Only 10% GPU Power

Problem Statement

Early versions of the Macaron platform required ~20 minutes to generate a mini‑app, exposing two fundamental issues in current AI agents: (1) reliance on Retrieval‑Augmented Generation (RAG) that stores isolated facts, and (2) lack of a persistent, habit‑aware memory.

Memory Diffusion Concept

In the technical report Exploring Agentic Memory , the Mind Lab team proposes treating memory as a policy rather than static storage. They introduce Memory Diffusion , a reinforcement‑learning (RL) framework that trains the model to both remember useful information and forget irrelevant data.

Scaling RL to a Trillion‑Parameter MoE Model

The target model is Kimi‑K2 , a 1.04 trillion‑parameter mixture‑of‑experts (MoE) model. Conventional full‑parameter RL would require a massive GPU fleet, but Mind Lab only had eight nodes equipped with 64 NVIDIA H800 GPUs.

Engineering Solution

Built on NVIDIA Megatron‑Bridge and added LoRA (Low‑Rank Adaptation) support.

Implemented zero‑copy data transfer between the inference engine ( vLLM) and the training engine ( Megatron) to avoid moving the full parameter set.

Introduced Truncated Importance Sampling to correct the policy lag caused by differing inference and training back‑ends, ensuring stable policy updates during fast inference.

Resource Efficiency

The combined architecture enabled RL training on the trillion‑parameter model while consuming only about 10 % of the GPU resources required by traditional full‑parameter RL pipelines.

Empirical Results

RL‑enhanced memory reduced Macaron’s app generation latency from ~20 minutes to ~2 minutes (≈10× speed‑up).

Benchmark comparing 1.5 B full‑parameter RL vs. 32 B LoRA‑based RL showed that, under identical compute budgets, the larger model with LoRA dramatically outperformed the smaller model trained end‑to‑end.

The codebase was merged into the main branches of both NVIDIA Megatron and ByteDance’s open‑source RL framework verl , indicating industry‑level validation.

Key Takeaway

Reinforcement learning’s performance ceiling is governed more by the pre‑trained model’s prior knowledge than by the amount of RL fine‑tuning. Leveraging a massive MoE model with lightweight LoRA adapters can achieve superior results with a fraction of the compute cost, providing a practical path for startups to build competitive, low‑latency AI products without relying on external API upgrades.