Research map / Large language models / Retrieval and knowledge

From RAG to Memory: Non-Parametric Continual Learning for Large Language Models

ICML 2025 · Large language models › Retrieval and knowledge · approach: RAG robustness and evaluation · 2 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes HippoRAG 2, a non-parametric continual-learning RAG framework that combines an open KG with passage nodes, query-to-triple linking, LLM triple filtering, and Personalized PageRank to improve factual, sense-making, and associative memory over standard RAG.

Problem

Standard vector-retrieval RAG cannot capture associativity (multi-hop connections) or sense-making over large contexts, while existing structure-augmented RAG methods degrade on simple factual memory tasks.

Approach

Extends HippoRAG's KG + Personalized PageRank by adding passage nodes to the graph (dense-sparse integration), linking queries to triples instead of NER entities (deeper contextualization), and using an LLM to filter retrieved triples (recognition memory) before PPR seeding.

Key results

• HippoRAG 2 achieves average F1 of 59.8 with Llama-3.3-70B vs 57.0 for NV-Embed-v2, the strongest baseline, and improves over it on 2Wiki by 9.5 F1 points.
• Passage recall@5 averages 78.2 for HippoRAG 2 vs 73.4 for NV-Embed-v2, with 5.0 and 13.9 point recall gains on MuSiQue and 2Wiki.
• Ablations show query-to-triple linking improves recall@5 by 12.5% over NER-to-node, and removing passage nodes drops average recall from 87.1 to 81.0.

Measured effects reported in the paper

Surprising findings

• LightRAG scores extremely low across all benchmarks (average F1 6.6 with Llama-3.3-70B), far below even no-retrieval.
• Query-to-node linking performs worse than NER-to-node on average, despite being a more general query representation.

Limitations

• Relies on heavy LLM use for offline OpenIE and online triple filtering, adding cost relative to standard RAG.
• Triple filter precision leaves room for improvement; 18% of error-analysis samples end with zero triples after filtering.
• Only evaluated on sampled subsets (1,000 queries per dataset) and the continual learning experiment is a simulated corpus expansion.

Experimental setup

Datasets: NaturalQuestions PopQA MuSiQue 2WikiMultihopQA HotpotQA LV-Eval NarrativeQA

Benchmarks: NaturalQuestions PopQA MuSiQue 2WikiMultihopQA HotpotQA LV-Eval NarrativeQA

Metrics: F1 exact match passage recall@5 passage recall@2 token counts indexing time QA time per query GPU memory

Baselines: BM25 Contriever GTR GTE-Qwen2-7B-Instruct GritLM-7B NV-Embed-v2 RAPTOR GraphRAG LightRAG HippoRAG

Builds on: HippoRAG Personalized PageRank DSPy MIPROv2

Closest papers by meaning

• HippoRAG: Neurobiologically Inspired Long-Term Memory for Large Language Models (NeurIPS 2024 · similarity 0.92)
• Long-Context LLMs Meet RAG: Overcoming Challenges for Long Inputs in RAG (ICLR 2025 · similarity 0.89)
• Adapting to Non-Stationary Environments: Multi-Armed Bandit Enhanced Retrieval-Augmented Generation on Knowledge Graphs (AAAI 2025 · similarity 0.88)
• RAG-DDR: Optimizing Retrieval-Augmented Generation Using Differentiable Data Rewards (ICLR 2025 · similarity 0.88)
• Scalable Language Model with Generalized Continual Learning (ICLR 2024 · similarity 0.88)
• CMT: A Memory Compression Method for Continual Knowledge Learning of Large Language Models (AAAI 2025 · similarity 0.88)