From RAG to Memory: Non-Parametric Continual Learning for Large Language Models (2502.14802v2)

Abstract: Our ability to continuously acquire, organize, and leverage knowledge is a key feature of human intelligence that AI systems must approximate to unlock their full potential. Given the challenges in continual learning with LLMs, retrieval-augmented generation (RAG) has become the dominant way to introduce new information. However, its reliance on vector retrieval hinders its ability to mimic the dynamic and interconnected nature of human long-term memory. Recent RAG approaches augment vector embeddings with various structures like knowledge graphs to address some of these gaps, namely sense-making and associativity. However, their performance on more basic factual memory tasks drops considerably below standard RAG. We address this unintended deterioration and propose HippoRAG 2, a framework that outperforms standard RAG comprehensively on factual, sense-making, and associative memory tasks. HippoRAG 2 builds upon the Personalized PageRank algorithm used in HippoRAG and enhances it with deeper passage integration and more effective online use of an LLM. This combination pushes this RAG system closer to the effectiveness of human long-term memory, achieving a 7% improvement in associative memory tasks over the state-of-the-art embedding model while also exhibiting superior factual knowledge and sense-making memory capabilities. This work paves the way for non-parametric continual learning for LLMs. Code and data are available at https://github.com/OSU-NLP-Group/HippoRAG.

• The paper introduces HippoRAG 2, a non-parametric continual learning method that improves Large Language Models' (LLMs) ability to integrate and retain new information without modifying the base model.
• HippoRAG 2 builds a dynamic knowledge graph from text using LLM-extracted triples and employs Personalized PageRank to enhance retrieval by connecting contextual information and filtering irrelevant data.
• Experiments demonstrate that HippoRAG 2 significantly outperforms standard Retrieval-Augmented Generation (RAG) on factual memory, associativity, and sense-making tasks, showcasing improved memory and reasoning capabilities.

The paper "From RAG to Memory: Non-Parametric Continual Learning for LLMs" introduces a new approach, HippoRAG 2, to improve how LLMs continually learn and remember information without changing the original model. This is important because in the real world, knowledge is always expanding, and AI needs to keep up.

Here's a breakdown of the key ideas:

1. The Problem: Limitations of Current Methods

LLMs are powerful, but they struggle with continual learning for two main reasons:

• They have trouble absorbing new information.
• They tend to forget old information when learning new things, a problem called catastrophic forgetting.

RAG (Retrieval-Augmented Generation) has become a popular solution. Instead of changing the LLM itself, RAG uses a separate database of knowledge. When the LLM needs information, it retrieves relevant passages from this database and uses them to generate an answer. But standard RAG has limitations. It struggles with:

• Sense-making: Understanding complex or uncertain contexts.
• Associativity: Connecting different pieces of information to draw conclusions.

Some existing methods try to improve RAG by adding structures like knowledge graphs, but they often perform worse on basic memory tasks compared to standard RAG.

2. The Solution: HippoRAG 2

HippoRAG 2 aims to fix these problems by:

• Improving upon the original HippoRAG framework.
• Using a method called Personalized PageRank to retrieve relevant information from a knowledge graph.
• Integrating passages more deeply into the retrieval process.
• Using the LLM more effectively to filter out irrelevant information.

3. How HippoRAG 2 Works

HippoRAG 2 has two main stages:

• Offline Indexing: This is like creating a well-organized filing system for information.
  • An LLM extracts "triples" from passages of text. A triple is a simple statement consisting of two phrases (subject and object) and a relation between them (e.g., "Einstein, was born in, Germany").
  • These triples are used to build a knowledge graph, where the phrases are nodes and the relations are edges connecting the nodes.
  • The system detects synonyms (words with similar meanings) and links them in the knowledge graph.
  • The original passages are also included in the knowledge graph, linking them to the phrases they contain.
• Online Retrieval: This is like searching the filing system to find the information you need to answer a question.
  • The system links the question to relevant triples and passages in the knowledge graph.
  • A "recognition memory" component filters out irrelevant triples using an LLM.
  • The Personalized PageRank algorithm is used to find the most relevant passages in the knowledge graph, considering the connections between different pieces of information.
  • The retrieved passages are then used by the LLM to generate an answer to the question.

Personalized PageRank (PPR) algorithm is a way to measure the importance of nodes in a network. It starts with a set of "seed nodes" (nodes that are initially considered important) and then iteratively updates the importance scores of all nodes based on their connections to the seed nodes. The "personalization" aspect means that the algorithm is biased towards certain nodes (the seed nodes) based on the query.

4. Key Improvements in HippoRAG 2

• Dense-Sparse Integration: Combines concise conceptual information (phrases) with richer contextual information (passages) in the knowledge graph.
• Deeper Contextualization: Matches the entire query to triples in the knowledge graph, capturing more of the query's meaning.
• Recognition Memory: Filters out irrelevant information using an LLM, improving the accuracy of retrieval.

5. Experiments and Results

The researchers tested HippoRAG 2 on a variety of question-answering tasks that measured:

• Factual Memory: Recalling simple facts.
• Associativity: Connecting multiple facts to answer a question.
• Sense-Making: Understanding complex narratives.

The results showed that HippoRAG 2 outperformed standard RAG methods on all three types of tasks, demonstrating its ability to improve both memory and reasoning. The experiments also showed that HippoRAG 2 is robust, meaning that it performs well with different LLMs and retrieval methods.

6. Why This Matters

HippoRAG 2 is a step towards creating LLMs that can continually learn and remember information more like humans. This could lead to AI systems that are more helpful and reliable in a wide range of real-world applications, where knowledge is constantly evolving.

7. Technical Details

• Knowledge Graph: A graph is a way of representing relationships between things. In this case, the "things" are phrases and passages, and the relationships are things like "is related to" or "contains".
• Triples: A simple statement consisting of two entities and a relation between them.
• Personalized PageRank: An algorithm used to rank the importance of nodes in a graph based on their connections to other nodes and a bias towards certain "seed" nodes.
• Embeddings: Numerical representations of text that capture their meaning. They allow the system to compare the similarity of different pieces of text.

In the paper, the researchers leverage an LLM to extract triples from each passage and integrates them into a schema-less open KG, which refers to the subject or object of these triples phrase and the edge connecting them relation edge. Then, an encoder identifies synonyms by evaluating phrase pairs within the KG, detecting those with vector similarity above a predefined threshold, and adding a synonym edge between such a pair. Also, the PPR algorithm then performs context-based retrieval on the KG to provide the most relevant passages for the final QA task.

HippoRAG 2 uses the open-source Llama-3.3-70B-Instruct as both the extraction (NER and OpenIE) and triple filtering model, and nvidia/NV-Embed-v2 as the retriever.

8. Potential Impact

The researchers state that their work may have various societal implications, but they do not identify any concerns that warrant specific emphasis beyond those generally associated with LLMs and information retrieval systems.