Hindsight: Structured Agent Memory that Retains, Recalls, and Reflects
Open accessAbstract
Christopher Latimer, Nicolò Boschi, Andrew Neeser, Chris Bartholomew, Gaurav Srivastava, Xuan Wang, Naren Ramakrishnan. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 3: System Demonstrations). 2026.
AI snapshot
Written from the paper’s full text.
TL;DR
HINDSIGHT is a memory system for AI agents that organizes long-term memory into four networks (world, experience, observation, opinion) and exposes retain, recall, and reflect operations. It reports strong benchmark accuracy, including 83.6% on LongMemEval and 83.2% on LoCoMo with a 20B open-source backbone, and 91.4% on LongMemEval with Gemini-3 Pro.
Key findings
- HINDSIGHT with a 20B open-source backbone reaches 83.6% on LongMemEval and 83.2% on LoCoMo, outperforming full-context GPT-4o (60.2% LongMemEval) and prior memory systems such as Zep (71.2% LongMemEval, 75.1% LoCoMo) and Memobase (75.8% LoCoMo).
- Scaling the backbone improves results: OSS-120B reaches 89.0% on LongMemEval and 85.7% on LoCoMo; Gemini-3 Pro reaches 91.4% on LongMemEval and 89.6% on LoCoMo, with the LongMemEval 91.4% described as the best result across all systems including frontier-backed ones.
- Per-category LongMemEval gains over full-context GPT-4o are largest on long-horizon categories: temporal reasoning rises from 45.1% to 79.7% (20B) and 91.0% (Gemini-3); preference from 20.0% to 66.7% (20B) and 80.0% (Gemini-3); multi-session from 44.3% to 79.7% (20B) and 87.2% (Gemini-3) per Table 4.
- The system is open-source under MIT, available via `pip install hindsight-all` and Docker, with 13.3K GitHub stars and 763 forks to date, and in production use at Fortune 500 enterprises as stated in the abstract.
Methods
- System/demo paper: HINDSIGHT is a compound AI system with TEMPR for temporal entity memory priming retrieval (retain and recall) and CARA for coherent adaptive reasoning agents (reflect), backed by PostgreSQL with pgvector.
- Evaluation uses LongMemEval (500 questions, conversations up to 1.5M tokens) and LoCoMo (multi-session human conversations up to 35 sessions), reporting overall accuracy and per-category breakdowns against full-context GPT-4o, Zep, Mem0, A-Mem, Memobase, Supermemory, and Backboard.
- Recall runs four parallel retrieval channels—HNSW vector search, BM25 keyword matching, graph traversal, and temporal filtering—merged with Reciprocal Rank Fusion and refined by a cross-encoder reranker, with a token budget controlling retrieved context.
Limitations
- The paper states: HINDSIGHT relies on LLM calls for fact extraction, entity resolution, and opinion formation, so extraction errors can propagate; all evaluation was on English-language benchmarks; opinion evolution has not been validated through formal user studies; the system requires PostgreSQL with pgvector (though the embedded Python package bundles its own instance); and temporal parsing uses rules plus a lightweight T5 model that may miss highly ambiguous or culturally specific temporal references.
- Evident from the text: this demo paper summarizes key results from a fuller evaluation in Latimer et al. (2025), so full architecture and evaluation details are not contained here.
- Evident from Table 2: benchmark comparisons include systems with different backbones and some missing entries (e.g., full-context GPT-4o has no LoCoMo score), so cross-system comparisons are not uniformly controlled.
Model-written summary. Check it against the paper before relying on it.
Integrity signals
Registry
Cites
Statistics
Claims the paper makes
2 of 3 cited sources checked · 2 supported · 0 flagged
- Premisecitation§ 2
The four-network design is inspired by cognitive science models of human memory.
“Our four-network design is inspired by cognitive science models of human memory (Sumers et al., 2024; Shan et al., 2025).”
Quote in the paperNot a substantial claim1 of 2 sources checked- cites Sumers et al., 2024
Couldn’t be matched to a paper in Crossref, so it can’t be checked.
- cites Shan et al., 2025 source
- Findingown data§ Table 2
HINDSIGHT with a 20B open-source model achieves 83.2% accuracy on LoCoMo.
“HINDSIGHT (OSS-20B) 83.6 83.2”
Quote in the paperNot a substantial claim- Rests on an assumption: The benchmark results for HINDSIGHT and the compared systems on LongMemEval and LoCoMo are directly comparable and not confounded by differences in evaluation protocols, prompt engineering, or backbone model capabilities.
- Findingreasoning§ 4.2
For a bank with 10,000 memory units, recall completes in under 200ms excluding the backbone LLM call.
“For a bank with 10,000 memory units, recall completes in under 200ms excluding the backbone LLM call”
Quote in the paperSubstantial claim - Findingown data§ 6.1 Overall Results
HINDSIGHT's 91.4% accuracy on LongMemEval is the best result across all systems including those backed by frontier models.
“Scaling the backbone to 120B reaches 89.0%, and using Gemini-3 Pro for answer generation pushes accuracy to 91.4%, the best result across all systems including those backed by frontier models.”
Quote in the paperSubstantial claim - Findingown data§ Abstract
HINDSIGHT with Gemini-3 Pro achieves 91.4% accuracy on LongMemEval.
“with Gemini-3 Pro, LongMemEval accuracy reaches 91.4%.”
- Also stated in: § Table 2 · plain
“HINDSIGHT (Gemini-3) 91.4 89.6”
Quote in the paperSubstantial claim- Rests on an assumption: The benchmark results for HINDSIGHT and the compared systems on LongMemEval and LoCoMo are directly comparable and not confounded by differences in evaluation protocols, prompt engineering, or backbone model capabilities.
- Also stated in: § Table 2 · plain
- Findingown data§ Table 2
HINDSIGHT with a 20B open-source model achieves 83.6% accuracy on LongMemEval.
“HINDSIGHT (OSS-20B) 83.6 83.2”
Quote in the paperSubstantial claim- Rests on an assumption: The benchmark results for HINDSIGHT and the compared systems on LongMemEval and LoCoMo are directly comparable and not confounded by differences in evaluation protocols, prompt engineering, or backbone model capabilities.
- Findingown data§ 6.1 Overall Results
HINDSIGHT with Gemini-3 Pro for answer generation achieves 91.4% accuracy on LongMemEval.
“Scaling the backbone to 120B reaches 89.0%, and using Gemini-3 Pro for answer generation pushes accuracy to 91.4%, the best result across all systems including those backed by frontier models.”
Quote in the paperSubstantial claim - Findingown data§ 6.1 Overall Results
HINDSIGHT with a 120B backbone reaches 89.0% accuracy on LongMemEval.
“Scaling the backbone to 120B reaches 89.0%, and using Gemini-3 Pro for answer generation pushes accuracy to 91.4%, the best result across all systems including those backed by frontier models.”
Quote in the paperSubstantial claim - Interpretationreasoning§ 3.3
This produces agents whose beliefs evolve with evidence rather than remaining fixed or getting overwritten.
“This produces agents whose beliefs evolve with evidence rather than remaining fixed or getting overwritten.”
Quote in the paperSubstantial claim - Interpretationown data§ 6.2 Where the Architecture Helps Most
The memory architecture matters more than model size.
“These results illustrate that the memory architecture matters more than model size.”
Quote in the paperSubstantial claim - Interpretationreasoning§ 4.2
The token budget mechanism ensures recall latency remains stable as the bank grows.
“and the token budget mechanism ensures this remains stable as the bank grows.”
Quote in the paperSubstantial claim
- Classified as probably not a substantial claim. 22
- May combine two claims. 15
Probably not a substantial claim: may not pass S3 (load-bearing).
Premisecitation§ 3.2This is conceptually related to verbal reinforcement in Reflexion, but applied to persistent opinion state rather than episodic task feedback.
“This is conceptually related to verbal reinforcement in Reflexion (Shinn et al., 2023), but applied to persistent opinion state rather than episodic task feedback.”
Quote in the paperNot a substantial claim1 of 1 sources checked- cites Shinn et al., 2023 source
Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based). May combine two claims.
Findingnone given§ AbstractHINDSIGHT has 13.3K GitHub stars and 763 forks to date.
“with 13.3K GitHub stars and 763 forks to date”
Quote in the paperNot a substantial claimProbably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.
Findingnone given§ AbstractHINDSIGHT is open-source under the MIT license, available as a Python package (pip install hindsight-all) and Docker image.
“The system is open-source under the MIT license, available as a Python package (pip install hindsight-all) and Docker image”
- Also stated in: § 5 · plain
“HINDSIGHT is available as an installable package via pip install hindsight-all and as a Docker image”
Quote in the paperNot a substantial claim- Also stated in: § 5 · plain
May combine two claims.
Findingown data§ Table 2HINDSIGHT with Gemini-3 achieves 91.4% on LongMemEval and 89.6% on LoCoMo.
“HINDSIGHT (Gemini-3) 91.4 89.6”
Quote in the paperSubstantial claim- Rests on an assumption: The benchmark results for HINDSIGHT and the compared systems on LongMemEval and LoCoMo are directly comparable and not confounded by differences in evaluation protocols, prompt engineering, or backbone model capabilities.
May combine two claims.
Findingown data§ Table 2HINDSIGHT with a 20B open-source model outperforms full-context GPT-4o and all prior memory systems on LongMemEval and LoCoMo.
“HINDSIGHT with a 20B open backbone beats full-context GPT-4o and prior systems.”
Quote in the paperSubstantial claim- Rests on an assumption: The benchmark results for HINDSIGHT and the compared systems on LongMemEval and LoCoMo are directly comparable and not confounded by differences in evaluation protocols, prompt engineering, or backbone model capabilities.
Probably not a substantial claim: may not pass S4 (based). May combine two claims.
Findingnone given§ 1CARA produces preference-conditioned responses and maintains opinions that evolve as new evidence arrives.
“CARA (Coherent Adaptive Reasoning Agents) produces preference-conditioned responses and maintains opinions that evolve as new evidence arrives.”
Quote in the paperNot a substantial claimProbably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.
Findingnone given§ 4.4Facts are extracted in the original language, entities are stored in their native script, and reflect responses match the query language.
“Facts are extracted in the original language, entities are stored in their native script, and reflect responses match the query language.”
Quote in the paperNot a substantial claimMay combine two claims.
Findingown data§ Table 2HINDSIGHT with OSS-120B achieves 89.0% on LongMemEval and 85.7% on LoCoMo.
“HINDSIGHT (OSS-120B) 89.0 85.7”
Quote in the paperSubstantial claim- Rests on an assumption: The benchmark results for HINDSIGHT and the compared systems on LongMemEval and LoCoMo are directly comparable and not confounded by differences in evaluation protocols, prompt engineering, or backbone model capabilities.
Probably not a substantial claim: may not pass S3 (load-bearing).
Findingnone given§ 4.4HINDSIGHT detects the input language at retain time and preserves it through the pipeline.
“HINDSIGHT detects the input language at retain time and preserves it through the pipeline.”
Quote in the paperNot a substantial claimProbably not a substantial claim: may not pass S3 (load-bearing), S4 (based). May combine two claims.
Study design and conductnone given§ LimitationsHINDSIGHT's temporal parsing uses a combination of rules and a lightweight T5 model, which handles common date expressions well but may miss highly ambiguous or culturally specific temporal references.
“Temporal parsing uses a combination of rules and a lightweight T5 model, which handles common date expressions well but may miss highly ambiguous or culturally specific temporal references.”
Quote in the paperNot a substantial claimProbably not a substantial claim: may not pass S3 (load-bearing).
Study design and conductnone given§ LimitationsAll evaluation was conducted on English-language benchmarks (LongMemEval and LoCoMo).
“All evaluation was conducted on English-language benchmarks (LongMemEval and LoCoMo).”
Quote in the paperNot a substantial claimProbably not a substantial claim: may not pass S3 (load-bearing).
Premisenone given§ Computed: evidence_span , evidence_countDirectives are user-defined hard rules that are injected into prompts at reflect time, with a priority order and an active flag.
“Directives are user-defined hard rules that are injected into prompts at reflect time, with a priority order and an active flag:”
Quote in the paperNot a substantial claimProbably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.
Premisereasoning§ 6.4 Qualitative Case StudyAfter session 3, the consolidation engine refines the existing observation: the new evidence weakens the earlier belief and the freshness trend flips from stable to weakening, while the world fact is unchanged.
“After session 3, the consolidation engine refines the existing observation: the new evidence weakens the earlier belief and the freshness trend flips from stable to weakening, while the world fact is unchanged.”
Quote in the paperNot a substantial claimProbably not a substantial claim: may not pass S3 (load-bearing), S4 (based). May combine two claims.
Study design and conductnone given§ Ethics StatementAll evaluation was conducted on publicly available benchmarks, and no personal data was collected.
“All evaluation was conducted on publicly available benchmarks, and no personal data was collected.”
Quote in the paperNot a substantial claimProbably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.
Premiseown data§ 6.4 Qualitative Case StudyIn a three-session interaction, Retain in session 1 creates a world fact, an experience fact, and updates the observation for the user-entity.
“Retain in session 1 creates a world fact (team rewrite in Rust), an experience fact (user began learning Rust), and updates the observation for the user-entity.”
Quote in the paperNot a substantial claimProbably not a substantial claim: may not pass S3 (load-bearing), S4 (based).
Interpretationnone given§ Broader Impact StatementThe open-source release allows the community to audit, extend, and improve safety mechanisms.
“The open-source release allows the community to audit, extend, and improve safety mechanisms.”
Quote in the paperNot a substantial claimClassified as probably not a substantial claim.
Premisenone given§ Ethics StatementHINDSIGHT stores and organizes information from user conversations in persistent memory banks.
“HINDSIGHT stores and organizes information from user conversations in persistent memory banks.”
Quote in the paperSubstantial claimProbably not a substantial claim: may not pass S3 (load-bearing), S4 (based). May combine two claims.
Interpretationnone given§ 4.3Avoiding a separate vector store removes a class of consistency bugs and keeps deployment to a single connection string, backup, and monitoring target.
“Avoiding a separate vector store removes a class of consistency bugs and keeps deployment to a single connection string, backup, and monitoring target.”
Quote in the paperNot a substantial claimProbably not a substantial claim: may not pass S3 (load-bearing), S4 (based). May combine two claims.
Interpretationnone given§ Ethics StatementHINDSIGHT's opinion formation mechanism reflects patterns in the data it processes and may amplify biases present in the input.
“The opinion formation mechanism reflects patterns in the data it processes and may amplify biases present in the input.”
Quote in the paperNot a substantial claimProbably not a substantial claim: may not pass S3 (load-bearing).
Interpretationreasoning§ 5This shows the epistemic separation that the four-network design provides.
“This shows the epistemic separation that the four-network design provides.”
Quote in the paperNot a substantial claimMay combine two claims.
Interpretationreasoning§ 4.2Reflect adds one LLM generation call on top of recall, so end-to-end query latency is driven primarily by the backbone model's inference speed.
“Reflect adds one LLM generation call on top of recall, so end-to-end query latency is driven primarily by the backbone model's inference speed.”
Quote in the paperNot a substantial claimProbably not a substantial claim: may not pass S3 (load-bearing).
Interpretationnone given§ 5The interface shows which retrieval channels contributed to each result, along with RRF fusion scores and reranker outputs, making the multi-strategy recall pipeline transparent.
“The interface shows which retrieval channels contributed to each result, along with RRF fusion scores and reranker outputs, making the multi-strategy recall pipeline transparent.”
Quote in the paperNot a substantial claimProbably not a substantial claim: may not pass S3 (load-bearing).
Interpretationreasoning§ Ethics StatementHINDSIGHT's persistent memory raises privacy considerations, as the system retains personal details, preferences, and opinions over time.
“This raises privacy considerations, as the system retains personal details, preferences, and opinions over time.”
Quote in the paperNot a substantial claimProbably not a substantial claim: may not pass S3 (load-bearing), S4 (based). May combine two claims.
Interpretationnone given§ 4.5Each memory bank is isolated with its own four-network structure and behavioral profile, so the system serves multiple agents concurrently without cross-contamination.
“Each memory bank is isolated with its own four-network structure and behavioral profile, so the system serves multiple agents concurrently without cross-contamination.”
Quote in the paperNot a substantial claimProbably not a substantial claim: may not pass S3 (load-bearing).
Premiseown data§ Computed: evidence_span , evidence_countHINDSIGHT provides drop-in wrappers for existing LLM clients, so developers can add memory to an agent without changing their application code.
“HINDSIGHT provides drop-in wrappers for existing LLM clients, so developers can add memory to an agent without changing their application code.”
Quote in the paperNot a substantial claimProbably not a substantial claim: may not pass S3 (load-bearing).
Study design and conductown data§ LimitationsPerformance on other languages has not been tested.
“Performance on other languages has not been tested.”
Quote in the paperNot a substantial claim
The paper relies on this without support.
The benchmark results for HINDSIGHT and the compared systems on LongMemEval and LoCoMo are directly comparable and not confounded by differences in evaluation protocols, prompt engineering, or backbone model capabilities.
“HINDSIGHT with a 20B open backbone beats full-context GPT-4o and prior systems.”
Self-check
Percentages, sums and reported differences re-computed from the paper’s own numbers.
Each contradiction between two passages listed here was confirmed by an independent check.
20B model ties rather than outperforms GPT-4o in one category
MinorThe single-session assistant category is 94.6% for both the 20B model and full-context GPT-4o, contradicting the claim that the 20B model outperforms it on every category.
“HINDSIGHT with a 20B open-source back- bone outperforms full-context GPT-4o on every cate- gory.”
“Single-sess. (asst) 94.6 80.4 94.6 98.2 96.4”
User action is classified as an agent experience
MinorThe experience network is defined as recording the agent's own actions, but the case study calls the user's beginning to learn Rust an experience fact.
“an experience fact (user began learning Rust)”
“Experience network: first-person records of the agent's own actions”
PostgreSQL requirement conflicts with Oracle backend support
MinorA supported Oracle backend with full feature parity contradicts the limitation's statement that the system requires PostgreSQL with pgvector.
“An Oracle AI Database backend with full feature parity is also supported.”
“The system requires PostgreSQL with pgvector,”
Cites
Chain-of-Thought Prompting Elicits Reasoning in Large Language Models
Wei, J., Wang, X., Schuurmans, D., Bosma, M., Ichter, B., Xia, F., Chi, E., Le, Q., & Zhou, D. (2022) · arXiv · doi:10.48550/arxiv.2201.11903 · Cited by 4,263 · see citing works
Self-citationBillion-scale similarity search with GPUs
Johnson, J., Douze, M., & Jegou, H. (2021) · IEEE Transactions on Big Data · doi:10.1109/tbdata.2019.2921572 · Cited by 4,228 · cited 4× · see citing works
“Entity links connect memories about the same person or thing across distant sessions, enabling multi-hop discovery that pure vector search would miss (Johnson et al., 2021).”
InfluentialThe probabilistic relevance framework: BM25 and beyond
Robertson, S., & Zaragoza, H. (2009) · Foundations and Trends® in Information Retrieval · doi:10.1561/1500000019 · Cited by 3,374 · cited 2× · see citing works
“Our multi-strategy re-call pipeline combines HNSW-based vector search (Malkov and Yashunin, 2020), BM25 key-word retrieval (Robertson and Zaragoza, 2009), graph traversal, and temporal filtering.”
Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks
Lewis, P., Perez, E., Piktus, A., Petroni, F., Karpukhin, V., Goyal, N., Küttler, H., Lewis, M., Yih, W.-T., Rocktäschel, T., et al. (2020) · arXiv · doi:10.48550/arxiv.2005.11401 · Cited by 3,101 · see citing works
Generative Agents: Interactive Simulacra of Human Behavior
Park, J. S., O'Brien, J. C., Cai, C. J., Morris, M. R., Liang, P., & Bernstein, M. S. (2023) · arXiv · doi:10.48550/arxiv.2304.03442 · Cited by 2,173 · see citing works
Malkov, Y. A., & Yashunin, D. A. (2020) · IEEE Transactions on Pattern Analysis and Machine Intelligence · doi:10.1109/tpami.2018.2889473 · Cited by 1,904 · cited 3× · see citing works
“HNSW indexes (Malkov and Yashunin, 2020) support sub-linear vector search, while GIN indexes enable BM25 (Robert-son and Zaragoza, 2009) full-text retrieval, both within a single database without requiring external search infrastructure like Elasticsearch or dedicated vector stores such as FAISS…”
Reflexion: Language Agents with Verbal Reinforcement Learning
Shinn, N., Cassano, F., Berman, E., Gopinath, A., Narasimhan, K., & Yao, S. (2023) · arXiv · doi:10.48550/arxiv.2303.11366 · Cited by 1,087 · see citing works
Reciprocal rank fusion outperforms condorcet and individual rank learning methods
Cormack, G. V., Clarke, C. L. A., & Buettcher, S. (2009) · Proceedings of the 32nd international ACM SIGIR conference on Research and development in information retrieval · doi:10.1145/1571941.1572114 · Cited by 797 · cited 4× · see citing works
“Results are merged using Reciprocal Rank Fusion (RRF) (Cormack et al., 2009) and refined with a cross-encoder reranker (Nogueira and Cho, 2019).”
InfluentialNogueira, R., & Cho, K. (2019) · arXiv · doi:10.48550/arxiv.1901.04085 · Cited by 615 · cited 3× · see citing works
“Results are merged using Reciprocal Rank Fusion (RRF) (Cormack et al., 2009) and refined with a cross-encoder reranker (Nogueira and Cho, 2019).”
ReAct: Synergizing Reasoning and Acting in Language Models
Yao, S., Zhao, J., Yu, D., Du, N., Shafran, I., Narasimhan, K., & Cao, Y. (2022) · arXiv · doi:10.48550/arxiv.2210.03629 · Cited by 608 · cited 1× · see citing works
“Large language models have unlocked a new generation of AI agents that can plan, reason, and take actions in open-ended environments (Yao et al., 2023; Wei et al., 2022; Shinn et al., 2023).”
MemoryBank: Enhancing Large Language Models with Long-Term Memory
Zhong, W., Guo, L., Gao, Q., Ye, H., & Wang, Y. (2023) · arXiv · doi:10.48550/arxiv.2305.10250 · Cited by 253 · see citing works
Augmented Language Models: a Survey
Mialon, G., Dessì, R., Lomeli, M., Nalmpantis, C., Pasunuru, R., Raileanu, R., Rozière, B., Schick, T., Dwivedi-Yu, J., Celikyilmaz, A., et al. (2023) · arXiv · doi:10.48550/arxiv.2302.07842 · Cited by 160 · see citing works
A Survey on the Memory Mechanism of Large Language Model-based Agents
Zhang, Z., Dai, Q., Bo, X., Ma, C., Li, R., Chen, X., Zhu, J., Dong, Z., & Wen, J.-R. (2025) · ACM Transactions on Information Systems · doi:10.1145/3748302 · Cited by 118 · cited 2× · see citing works
“…surveys on LLM-based agent memory identify three core challenges that remain open: how to retain the right information from growing interaction histories, how to recall it efficiently at query time, and how to reason over it coher-ently (Zhang et al., 2025b; Wu et al., 2025; Zhang et al., 2025a).”
Cognitive Architectures for Language Agents
Sumers, T. R., Yao, S., Narasimhan, K., & Griffiths, T. L. (2023) · arXiv · doi:10.48550/arxiv.2309.02427 · Cited by 72 · see citing works
MemGPT: Towards LLMs as Operating Systems
Packer, C., Wooders, S., Lin, K., Fang, V., Patil, S. G., Stoica, I., & Gonzalez, J. E. (2023) · arXiv · doi:10.48550/arxiv.2310.08560 · Cited by 64 · cited 3× · see citing works
“…by research on cognitive memory in LLMs (Shan et al., 2025) and surveys mapping human memory structures to agent architectures (Wu et al., 2025) (Packer et al., 2023), which focuses on tiered context management, or A-Mem (Xu et al., 2025), which builds agentic memory without temporal reasoning,…”
A-Mem: Agentic Memory for LLM Agents
Xu, W., Liang, Z., Mei, K., Gao, H., Tan, J., & Zhang, Y. (2025) · Advances in Neural Information Processing Systems 38 · doi:10.52202/085713-0593 · Cited by 51 · cited 5× · see citing works
“…INDSIGHT with an open-source 20B model achieves 83.6% and 83.2% overall accuracy, outperforming full-context GPT-4o and prior memory systems such as Zep (Rasmussen et al., 2025), Mem0 (Chhikara et al., 2025), and A-Mem (Xu et al., 2025); scaling to Gemini-3 Pro pushes LongMemEval accuracy to 91.4%.”
InfluentialHindsight is 20/20: Building Agent Memory that Retains, Recalls, and ReflectsPreprint
Latimer, C., Boschi, N., Neeser, A., et al. (2025) · arXiv.org · doi:10.48550/arxiv.2512.12818 · Cited by 37 · cited 3× · see citing works
“The full architecture is in (Latimer et al., 2025); here we focus on system design and demonstration.”
InfluentialSelf-citationKARMA: Augmenting Embodied AI Agents with Long-and-Short Term Memory Systems
Wang, Z., Yu, B., Zhao, J., Sun, W., Hou, S., Liang, S., Hu, X., Han, Y., & Gan, Y. (2025) · 2025 IEEE International Conference on Robotics and Automation (ICRA) · doi:10.1109/icra55743.2025.11128047 · Cited by 11 · cited 2× · see citing works
“KARMA (Wang et al., 2025) applies long-and-short term memory to embodied agents, while MemVerse (Liu et al., 2025) extends memory to multimodal lifelong learning settings.”
Mem0: Building Production-Ready AI Agents with Scalable Long-Term Memory
Chhikara, P., Khant, D., Aryan, S., Singh, T., & Yadav, D. (2025) · arXiv · doi:10.48550/arxiv.2504.19413 · Cited by 10 · see citing works
Zep: A Temporal Knowledge Graph Architecture for Agent Memory
Rasmussen, P., Paliychuk, P., Beauvais, T., Ryan, J., & Chalef, D. (2025) · arXiv · doi:10.48550/arxiv.2501.13956 · Cited by 9 · cited 4× · see citing works
“Zep (Rasmussen et al., 2025) builds temporal knowledge graphs from conversations and supports time-aware retrieval.”
InfluentialOn the Failure of Latent State Persistence in Large Language Models
Huang, J.-T., Sun, K., Wang, W.-X., et al. (2025) · doi:10.48550/arxiv.2505.10571 · Cited by 6
Evaluating Very Long-Term Conversational Memory of LLM Agents
Maharana, A., Lee, D.-H., Tulyakov, S., Bansal, M., Barbieri, F., & Fang, Y. (2024) · arXiv · doi:10.48550/arxiv.2402.17753 · Cited by 3 · cited 6× · see citing works
“We test on two established benchmarks for long-term conversational memory: LongMemEval (Wu et al., 2024), which contains 500 questions across conversations spanning up to 1.5M tokens, and LoCoMo (Maharana et al., 2024), which uses multi-session human conversations with up to 35 sessions each.”
InfluentialLongMemEval: Benchmarking Chat Assistants on Long-Term Interactive Memory
Wu, D., Wang, H., Yu, W., Zhang, Y., Chang, K.-W., & Yu, D. (2024) · arXiv · doi:10.48550/arxiv.2410.10813 · Cited by 3 · cited 3× · see citing works
“We test on two established benchmarks for long-term conversational memory: LongMemEval (Wu et al., 2024), which contains 500 questions across conversations spanning up to 1.5M tokens, and LoCoMo (Maharana et al., 2024), which uses multi-session human conversations with up to 35 sessions each.”
InfluentialFrom Human Memory to AI Memory: A Survey on Memory Mechanisms in the Era of LLMs
Wu, Y., Liang, S., Zhang, C., Wang, Y., Zhang, Y., Guo, H., Tang, R., & Liu, Y. (2025) · arXiv · doi:10.48550/arxiv.2504.15965 · Cited by 3 · cited 2× · see citing works
“…surveys on LLM-based agent memory identify three core challenges that remain open: how to retain the right information from growing interaction histories, how to recall it efficiently at query time, and how to reason over it coher-ently (Zhang et al., 2025b; Wu et al., 2025; Zhang et al., 2025a).”
Memory in Large Language Models: Mechanisms, Evaluation and EvolutionPreprint
Zhang, D., Li, W., Song, K., Lü, J., Li, G., Yang, L., & Li, S. (2025) · arXiv (Cornell University) · doi:10.48550/arxiv.2509.18868 · Cited by 2 · cited 3× · see citing works
“…surveys on LLM-based agent memory identify three core challenges that remain open: how to retain the right information from growing interaction histories, how to recall it efficiently at query time, and how to reason over it coher-ently (Zhang et al., 2025b; Wu et al., 2025; Zhang et al., 2025a).”
Yan, S., Yang, X., Huang, Z., Nie, E., Ding, Z., Li, Z., Ma, X., Bi, J., Kersting, K., Pan, J. Z., et al. (2025) · arXiv · doi:10.48550/arxiv.2508.19828 · Cited by 2 · see citing works
Cognitive Memory in Large Language Models
Shan, L., Luo, S., Zhu, Z., Yuan, Y., & Wu, Y. (2025) · arXiv · doi:10.48550/arxiv.2504.02441 · Cited by 1 · see citing works
Beyond a Million Tokens: Benchmarking and Enhancing Long-Term Memory in LLMs
Tavakoli, M., Salemi, A., Ye, C., Abdalla, M., Zamani, H., & Mitchell, J. R. (2025) · arXiv · doi:10.48550/arxiv.2510.27246 · cited 3×
“This component also supports Mem-oryBench (Ai et al., 2025) questions, providing additional coverage of memory and continual learning evaluation scenarios (Tavakoli et al., 2025).”
MemoryBench: A Benchmark for Memory and Continual Learning in LLM Systems
Ai, Q., Tang, Y., Wang, C., Long, J., Su, W., & Liu, Y. (2025) · arXiv · doi:10.48550/arxiv.2510.17281 · cited 3×
“This component also supports Mem-oryBench (Ai et al., 2025) questions, providing additional coverage of memory and continual learning evaluation scenarios (Tavakoli et al., 2025).”
MemVerse: Multimodal Memory for Lifelong Learning Agents
Liu, J., Sun, Y., Cheng, W., Lei, H., Chen, Y., Wen, L., Yang, X., Fu, D., Cai, P., Deng, N., et al. (2025) · arXiv · doi:10.48550/arxiv.2512.03627
Flagged first, then by citations
Cited by
Sorted by citations
Remember by Asking: Retrieval-Induced Memory Evolution for LLM AgentsPreprint
Zhou, W., Lu, J., Wang, Y., Xing, Z., Chen, Z., Han, A., & Shi, H. (2026) · arXiv (Cornell University) · doi:10.48550/arxiv.2609.34438
“Hind-sight [13] combines multiple retrieval strategies with explicit reflection over structured memories, while SAGE [14] and REALM [15] use retrieval feedback to further evolve or re-consolidate the memory structure.”
TRACE: Governing Memory Validity in Evolving Multi-Agent Systems
Xiong, W.-J., Zhang, S., Gu, S., et al. (2026) · doi:10.48550/arxiv.2609.33517
“…or skills (Packer et al., 2023; Xu et al., 2025), while newer managers organize, summarize, or time-index long-term conversations (Kang et al., 2025; Salama et al., 2025; Tan et al., 2025b; Ong et al., 2025; Banerjee et al., 2026; Li et al., 2026a; Ye et al., 2026; Latimer et al., 2026).”
Shared Worlds, Private Minds: Structured Memory for Long-Form Writing as World Creation
Tian, Q.-Y., Gu, X.-W., Su, H., et al. (2026) · doi:10.48550/arxiv.2609.32401
“Representative query-centric baselines span flat retained-memory stores (Mem0), temporal or entity-event graphs (Graphiti and AriGraph), and typed or multi-store organizations (Hindsight and MemoryOS) (Chhikara et al., 2025; Rasmussen et al., 2025; Latimer et al., 2026; Anokhin et al., 2025; Kang et al., 2025).”
Machine Unlearning for Large Language Models: Foundations, Advances, and Agentic Extensions
Xu, X., Du, M.-X., Bai, L., et al. (2026) · doi:10.48550/arxiv.2609.30909
“Functional gains require matched controls and do not establish target removal.”
- Authors
- Latimer, C., Boschi, N., Neeser, A., Bartholomew, C., Srivastava, G., Wang, X., & Ramakrishnan, N.
- Affiliations
- Virginia Tech (US)
- ORCID
- Latimer, Christopher
- Venue
- Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 3: System Demonstrations)
- Published
- 2026
- Pages
- 275-285
- OpenAlex
- W7166866500
- Cited by
- 0
- References
- 28
- FWCI
- 0.00
- Type
- conference paper
- Open access
- hybrid
- doi.org
- Licence
- cc-by
- Locations
- Annual Meeting of the Association for Computational Linguistics (ACL) (publishedVersion, cc-by)
- Topics
- Physical Sciences › Computer Science › Artificial Intelligence › AI-based Problem Solving and Planning
- Physical Sciences › Computer Science › Artificial Intelligence › Natural Language Processing Techniques
- Physical Sciences › Computer Science › Artificial Intelligence › Speech and dialogue systems
- Keywords
- memory retention, memory retrieval, reflection, large language models
About these checks
Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.
Claims read by deepseek-v4.1-flash under definition version 3 (extract-v3) on 2026-10-08.
Classified by jev-1.13-20260917 (gold-guideline-v0.2) on 2026-10-08.
Cited sources are checked by gpt-6-sol.
Statistics and arithmetic are recomputed, not judged.
Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.
What counts as a claim ›How we measure confidence ›How we check this ›
Are you an author, or do you see a problem with this page? Request a review
