HippoRAG: Neurobiologically Inspired Long-Term Memory for Large Language Models

In order to thrive in hostile and ever-changing natural environments, mammalian brains evolved to store large amounts of knowledge about the world and continually integrate new information while avoiding catastrophic forgetting. Despite the impressive accomplishments, large language models (LLMs), even with retrieval-augmented generation (RAG), still struggle to efficiently and effectively integrate a large amount of new experiences after pre-training. In this work, we introduce HippoRAG, a novel retrieval framework inspired by the hippocampal indexing theory of human long-term memory to enable deeper and more efficient knowledge integration over new experiences. HippoRAG synergistically orchestrates LLMs, knowledge graphs, and the Personalized PageRank algorithm to mimic the different roles of neocortex and hippocampus in human memory. We compare HippoRAG with existing RAG methods on multi-hop question answering and show that our method outperforms the state-of-the-art methods remarkably, by up to 20%. Single-step retrieval with HippoRAG achieves comparable or better performance than iterative retrieval like IRCoT while being 10-30 times cheaper and 6-13 times faster, and integrating HippoRAG into IRCoT brings further substantial gains. Finally, we show that our method can tackle new types of scenarios that are out of reach of existing methods. Code and data are available at https://github.com/OSU-NLP-Group/HippoRAG.

Paper

Similar papers

Peer review

Reviewer oPeR7/10 · confidence 4/52024-07-04

Summary

This paper proposes a retrieval-augmented generation (RAG) method that is inspired by the hippocampal indexing theory of human memory to enable longer knowledge storage and efficient knowledge integration over new experiences.

Strengths

1. This paper's idea is interesting and shows impressive performances. 2. The paper is well-written and the presentation is clear. 3. The metric `Node Specificity` is aligned with the intuition that humans may get better memorization of things that they are seeing over and over again. This is shown in Figure 2 where the logo of "Stanford" grows larger. I think this one is very interesting.

Weaknesses

1. Extracting triplets from the passage strongly depends on the power of the triplets extracting model, which is a language model in the paper's implementation. I'm concerned that it may lose information when the passage becomes longer. Is this a common strategy in other retrieval methods? 2. Missing citations. Since the paper talks about long-term memory in related work, I believe the following papers may need to be cited. [1] Memoria: Resolving Fateful Forgetting Problem through Human-Inspired Memory Architecture. [2] MEMORYLLM: Towards Self-Updatable Large Language Models. [3] CAMELoT: Towards Large Language Models with Training-Free Consolidated Associative Memory.

Questions

see weaknesses

Rating

7

Confidence

4

Soundness

3

Presentation

3

Contribution

3

Limitations

yes, the authors have addressed the limitations.

Reviewer bGvw5/10 · confidence 4/52024-07-10

Summary

This paper presents Hippo-RAG, which enables knowledge integration across retrieval results and supports long-term memory with a mechanism that resembles the hippocampal memory indexing theory. Hippo-RAG includes two steps: offline indexing to extract, encode, and index the passages to KG, and online retrieval to extract entities from the query and retrieve the results from the KG. Hippo-RAG outperforms various existing methods on the multi-hop QA benchmarks and is able to perform multi-hop retrieval in a single step.

Strengths

1. The proposed method is novel, and it resembles the long-term memory mechanism of human beings. The offline indexing stage enables knowledge integration across passages and addresses the path-finding problems in RAG. The pipeline examples can help to better understand the process. 2. The proposed method achieves significant performance improvement from various challenging multi-hop QA benchmarks. 3. The paper provides sufficient details for reproduction.

Weaknesses

1. The mechanism is great, but the generalization of the method is not good enough. Intuitively, the mentioned mechanism should work for several scenarios that require long-term memory. Still, the proposed method narrows it down and limits the experiments to multi-hop reasoning. 2. The framework's performance depends on the NER ability of LLM. However, [1] demonstrates a large performance gap between prompting LLMs for NER and fine-tuned NER models. The error analysis also indicates that most errors in the system are from NER. 3. The comparison of baselines may not be fair since the parameter sizes of Hippo-RAG are significantly larger than other methods. Contriever and ColBERTv2 achieve comparable results on MuSiQue and HotpotQA, but when combined with HippoRAG, the improvement is insignificant. ColBERTv2 is the second-best model, and the improvement from Hippo-RAG seems incremental on the two datasets. [1] https://arxiv.org/abs/2304.10428

Questions

1. Can Hippo-RAG work for other tasks that require long-term memory? 2. Why does Hippo-RAG perform significantly better on 2Wiki?

Rating

5

Confidence

4

Soundness

3

Presentation

3

Contribution

2

Limitations

The authors adequately discussed the limitations and potential negative societal impact of their work.

Authorsrebuttal2024-08-07

Rebuttal by Authors (Continued)

- **W3: The comparison of baselines may not be fair since the parameter sizes of Hippo-RAG are significantly larger than other methods. Contriever and ColBERTv2 achieve comparable results on MuSiQue and HotpotQA, but when combined with HippoRAG, the improvement is insignificant. ColBERTv2 is the second-best model, and the improvement from Hippo-RAG seems incremental on the two datasets.** In the response below, we present evidence and arguments that we hope will convince the reviewer that our performance improvements are substantial and our experimental setting is sound and fair. #### **Improvements on MuSiQue** We first highlight that our method “achieves significant performance improvement from various challenging multi-hop QA benchmarks”, as pointed out by the reviewer when listing our strengths. HippoRAG improves R@5 on MuSiQue by 5.5% and 2.7% over Contriever and ColBERTv2 respectively, as well as 3.4% F1 score on QA compared to ColBERTv2. Many well-received and highly cited works such as IRCoT [2], Self-Ask [3], ITER-RETGEN [4] and MCR [5] highlight improvements similar to ours. #### **HotpotQA’s Weaknesses as a Knowledge-Integration Benchmark** As discussed in subsection “Single-Step Retrieval Results” (Section 4) of our paper, our lower performance on HotpotQA is mainly due to its lower need for knowledge integration due to existing shortcut signals. This HotpotQA limitation is referenced when constructing MuSiQue [6], but also explored more deeply in Appendix B of our paper. In bridge multi-hop questions, the query and second supporting passage should be linked only through a bridge entity. However, as shown in Figure 6, HotpotQA queries are on average as similar to the second supporting passage as to their distractors, instead of less similar like in the other datasets, making that second supporting document easier to detect without knowledge integration. #### **Measuring the Impact of Improved Knowledge-Integration Directly** Additionally, when we measure the impact of improved knowledge integration in MuSiQue and 2WikiMultiHopQA, our improvements are even larger than our overall results. In Table 8, we show that HippoRAG’s ability to find all supporting documents is key to its strong performance, obtaining a 6.3% and 38.6% improvement over ColBERTv2 in All-Recall@5, compared to a 2.7% and 21.3% improvement in standard Recall@5. #### **Our Experimental Setting is Sound and Fair** Finally, we refer to the reviewer’s concern that our comparison with Contriever and ColBERTv2 might not be fair given their small size compared to LLM. We understand the reviewer’s apprehension, however, we argue that improving the performance of a system using an LLM and comparing it to the original system is a well-accepted paradigm in AI research. As a specific example, many well-received QA works such as IRCoT [2], Self-Ask [3] and RAPTOR [7] leverage a large LLM in different ways and compare with the original pipeline; which in these papers contain various retrieval methods like BM25, DPR [8] and ColBERTv2. We note that HippoRAG follows this same research paradigm, which is related to the large and growing body of recent work that leverages LLMs to challenge prior state of the art based on smaller models. As a final remark, we also note that our method outperforms RAPTOR and is comparable with IRCoT (and considerably more efficient). Since both of these methods augment the original retrieval process with outputs from an LLM, they can be seen as HippoRAG’s most direct baselines in that sense. - **Q2: Why does Hippo-RAG perform significantly better on 2Wiki?** As discussed in subsection “Single-Step Retrieval Results” (Section 4) of our paper, 2WikiMultiHopQA’s construction is more entity-centric than the other two datasets, making it particularly well-suited for HippoRAG’s design. To provide some extra context, 2WikiMultiHopQA was created by leveraging the Wikidata KG to determine which entities and relations could be found in Wikipedia passages to create compositional, comparison, inference or bridge-comparison multi-hop questions. Due to this construction process, queries in this dataset include at least one named entity, a characteristic that our methodology can leverage given its previously mentioned reliance on NER to link queries to the KG.

Authorsrebuttal2024-08-07

Rebuttal by Authors (Continued)

### References [1] Zhou et al. (2024). UniversalNER: Targeted Distillation from Large Language Models for Open Named Entity Recognition.\ [2] Trivedi et al. (2023). Interleaving Retrieval with Chain-of-Thought Reasoning for Knowledge-Intensive Multi-Step Questions.\ [3] Press et al. (2023). Measuring and Narrowing the Compositionality Gap in Language Models.\ [4] Shao et al. (2023). Enhancing Retrieval-Augmented Large Language Models with Iterative Retrieval-Generation Synergy.\ [5] Yoran et al. (2023). Answering Questions by Meta-Reasoning over Multiple Chains of Thought.\ [6] Trivedi et al. (2022). ♫ MuSiQue: Multihop Questions via Single-hop Question Composition.\ [7] Sarthi et al. (2024). RAPTOR: Recursive Abstractive Processing for Tree-Organized Retrieval.\ [8] Karpukhin et al. (2020). Dense Passage Retrieval for Open-Domain Question Answering.

Reviewer fKDL7/10 · confidence 4/52024-07-13

Summary

The paper introduces HippoRAG, a retrieval framework inspired by hippocampal indexing theory to enhance large language models (LLMs) in integrating new information. The algorithm is a combination of LLMs, knowledge graphs (KGs), and the Personalized PageRank algorithm. HippoRAG outperforms existing retrieval-augmented generation (RAG) methods in multi-hop question answering (QA) by a significant margin. It achieves comparable or better performance than iterative retrieval methods like IRCoT while being significantly cheaper and faster.

Strengths

- The idea of converting the corpus to KG and then running page rank algorithm for better retrievals is interesting. The connections to hippocampal memory theory is inspiring. - The method is shown to be more efficient than iterative retrieval method IRCoT and can also use IRCoT to further boost the performance. - The experiment is well-executed, showing the improvements of HippoRAG. - The paper is well-written and easy to follow. Analysis is clear.

Weaknesses

- Some baselines from KG-LLM for multi-hop QA literature are missing. These could be beneficial to replace Page Rank for ablation study or comparing with KGQA on open-source knowledge graph. - Some citation are missing (see questions)

Questions

- Can we evaluate the performance of constructed knowledge graph alone? - Some citations are missing: - Park et al. Graph Elicitation for Guiding Multi-Step Reasoning in Large Language Models - Jin et al. Improving embedded knowledge graph multi-hop question answering by introducing relational chain reasoning

Rating

7

Confidence

4

Soundness

3

Presentation

4

Contribution

3

Limitations

Yes

Reviewer fKDL2024-08-10

Thank you!

Thank the authors for their detailed rebuttal. My concerns are addressed adequately and I will keep my score.

Reviewer bGvw2024-08-11

Thank you for the analyses provided. I raise the rating from 4 to 5 for they address most of my concerns. My remaining concerns are about W3 experimental comparison. The settings for RAPTOR are not stated in the paper (which retriever (SBERT/DPR/BM25) is used for its implementation). In Table 2, ColBERTv2 outperforms the two LLM-augmented methods, and it is not discussed in the paper. These are my concerns, but they are not the weakness of the paper. I would appreciate it if the authors could include them in later versions.

Authorsrebuttal2024-08-12

Thank you for considering our responses and kindly updating your score. Please feel free to let us know if you have any further questions or concerns. We appreciate your suggestion to include more details about the LLM-augmented baselines. In order to ensure our experiments appropriately represent these baselines, we used the embedding methods which obtain the strongest performance in their respective experiments. Both of these were Sentence Transformer models, `sentence-transformers/multi-qa-mpnet-base-cos-v1` for RAPTOR and GTR (`sentence-transformers/gtr-t5-base`) for the Proposition-izer. In order to provide a direct comparison with ColBERTv2 as the reviewer suggests, we ran these two baselines using ColBERTv2 as their retrieval component. | Method | MuSiQue | MuSiQue | 2Wiki | 2Wiki | HotpotQA | HotpotQA | |---------------------------|----------|----------|---------|---------|----------|----------| | | R@2 | R@5 | R@2 | R@5 | R@2 | R@5 | | ColBERTv2 | 37.9 | 49.2 | 59.2 | 68.2 | **64.7** | **79.3** | | RAPTOR | 35.7 | 45.3 | 46.3 | 53.8 | 58.1 | 71.2 | | RAPTOR (ColBERTv2) | 36.9 | 46.5 | 57.3 | 64.7 | 63.1 | 75.6 | | Proposition | 37.6 | 49.3 | 56.4 | 63.1 | 58.7 | 71.1 | | Proposition (ColBERTv2) | 37.8 | 50.1 | 55.9 | 64.9 | 63.9 | 78.1 | | HippoRAG (ColBERTv2) | **40.9** | **51.9** | **70.7**| **89.1**| 60.5 | 77.7 | We find that both baselines obtain stronger performance using ColBERTv2 than their original models. However, they still underperform ColBERTv2 itself in all results except for MuSiQue (R@5), where only the Proposition-izer model outperforms ColBERTv2 and gets 0.8% closer to HippoRAG. We point out that RAPTOR's poor performance compared to ColBERTv2 demonstrates that their "cluster and summarize" methodology is mostly ineffective for the comprehensive knowledge integration required in these datasets. Finally, we note that the Proposition-izer's unstable performance, especially its poor performance in 2Wiki, illustrates that separating passages based on propositions can sharply diminish a retriever's knowledge integration capabilities. We will include these results and discussion in the revised version.

Reviewer oPeR2024-08-13

Response to the rebuttal

Thank the authors for the responses. I don't have more concerns.

Program Chairsdecision2024-09-25

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC