Intelligence Begins with Memory

Open-Source Memory Operating System

A native memory framework developed by the MemTensor team for building intelligent systems capable of remembering, adapting, and evolving.

LightMem

Lightweight Pluggable Memory System for LLM

Addressing common challenges faced by large models in long-term interactions—such as limited context, information forgetting, and memory redundancy—LightMem introduces a human-memory-inspired lightweight architecture. Through semantic compression and noise filtering, it significantly reduces redundant information; dynamic semantic segmentation improves memory organization and retrieval efficiency; and an offline “sleep-like” update mechanism enables memory reorganization and knowledge consolidation. LightMem maintains high memory fidelity and consistency while reducing computational and storage costs, providing an efficient and scalable long-term memory solution for large language models and intelligent agents, supporting personalized and complex application scenarios.

Text2Mem

Unified Memory Operation Language

Proposes a unified language, Text2Mem, for intelligent agent memory management. It converts natural language into standardized JSON commands covering various operations such as encoding, storage, and retrieval. Through parsing, validation, and adaptation layers, it enables secure, deterministic, and portable execution across heterogeneous memory backends, providing a standardized foundation for memory control.

KaLM-V2

High-Performance Universal Text Encoder

To address the balance challenge between retrieval efficiency and precision in RAG systems, a multi-scale universal embedding model is proposed. In the MTEB benchmark—which covers thousands of languages, hundreds of tasks, and nine major task types—it ranks first globally, surpassing major open- and closed-source models such as NVIDIA’s nvidia/llama-embed-nemotron-8b, Alibaba’s Qwen/Qwen3-Embedding-8B, and Google’s Google/gemini-embedding-001. It has been widely deployed and downloaded over one million times.

MoM

Document Memory Extraction for RAG

Upgrades traditional RAG’s passive chunking into proactive document memory construction. It leverages expert-level large models to generate logical outlines and key segments, combines multi-path sampling and multi-dimensional evaluation to select high-quality memories, and trains smaller models to acquire human-like reading and three-tier memory retrieval capabilities—enhancing multi-domain retrieval and generation performance.

MoC

Hybrid Document Chunking Expert for RAG

Introduces two new metrics—boundary clarity and block cohesion—for evaluating text segmentation quality, and builds a granularity-aware hybrid chunking framework. This guides models to generate regularized segmentation rules, producing structured, high-quality text chunks that significantly improve RAG retrieval and generation performance while maintaining a balance between computational efficiency and precision.

HaluMem

Memory Hallucination Evaluation Framework

Proposes the industry’s first hallucination evaluation framework designed for memory operating systems. It supports fine-grained and procedural hallucination assessment, helping memory systems quickly locate hallucination types and identify optimization directions.

RecCocktail

Parameterized Memory Fusion Method for Personalized Recommendation Systems

For personalized recommendation scenarios, this method adaptively fuses domain-general user parameterized memory with scenario-specific user parameterized memory in the weight space. Through a plug-and-play modular design, it enables efficient integration of any domain memory into core recommendation capabilities without additional inference overhead. It significantly improves personalized recommendation model performance across diverse application scenarios.

SafeRAG

RAG Security Evaluation Benchmark

SafeRAG SafeRAG introduces a security evaluation benchmark for RAG systems, constructing datasets that cover various attack tasks such as noise injection, context conflict, soft advertising, and denial-of-service. The benchmark systematically simulates multiple attack scenarios and empirically reveals critical vulnerabilities of existing RAG components when facing malicious knowledge manipulation.

CRUD-RAG

Comprehensive Chinese RAG Evaluation Benchmark

Builds a large-scale Chinese retrieval-augmented generation evaluation benchmark. It systematically designs datasets and metrics for four core application scenarios—Create, Read, Update, and Delete—jointly evaluating the performance of retrievers, knowledge bases, and large models. It provides empirical references for optimizing RAG systems under different real-world conditions.