The DOI matches its Crossref record; 28 of the 28 references it cites were found in a registry or an index; 23 of the 29 sources its claims cite were checked for support.
From Goldfish to Elephant: Long-Term Memory for Large Language Models via Selective Storage and Hierarchical Compression
Open accessAbstract
Large language models (LLMs) demonstrate remarkable reasoning and generative capabilities but remain fundamentally limited by fixed context windows and the absence of persistent, structured memory. As a result, they struggle with long-term personalization, continual learning, and consistency across extended interactions. In this paper, we propose a modular memory-augmented architecture that integrates episodic and semantic memory with a consistency-aware retrieval mechanism and a learned memory-gating network. Episodic memory stores recent, high-fidelity interaction traces, while semantic memory maintains compressed, abstracted knowledge derived from historical experiences. A hierarchical compression module regulates memory growth by summarizing episodic content into semantic representations when capacity thresholds are exceeded. To retrieve relevant information, we introduce a consistency-aware scoring function that jointly optimizes semantic similarity, temporal relevance, and contradiction avoidance. This retrieval output is fused into the input context of a transformer-based LLM to support grounded and coherent generation. A lightweight gating network predicts the utility of storing new interactions, enabling selective memory updates and reducing catastrophic interference during continual learning. We show that this design supports long-horizon dialogue coherence, improved factual consistency, and adaptive personalization without requiring model retraining. The proposed framework is model-agnostic and can be integrated with existing retrieval-augmented generation pipelines. Our results highlight a practical pathway toward scalable long-term memory for LLMs, bridging the gap between stateless neural generation and persistent, structured knowledge systems.
Overview
CoChat Summary
TL;DR
MEMLLM is a memory-augmented LLM architecture combining selective gating, episodic/semantic hierarchical compression, consistency-aware retrieval, and elastic weight consolidation for long-term personalization. It reports 94.3% personalization accuracy, 73% memory reduction, and 68% fewer contradictions versus RAG baselines over 10,000+ interactions.
Key findings
- Selective gating achieves 0.893 F1, 91.2% precision, 87.4% recall, and 27.3% selectivity, storing only about 1 in 4 interactions; without gating, accuracy drops to 89.1% and memory rises to 647.2 MB.
- Hierarchical compression at ρ = 0.2 gives 91.3% reconstruction F1 and 5× size reduction; without compression, memory grows to 891.4 MB, and memory stabilizes around 1096 total entries (519 episodic + 577 semantic).
- Consistency-aware retrieval reduces contradiction rate to 6.3% versus 28.2% for RAG-Dense; without consistency scoring, consistency falls to 79.4%.
- Continual learning with EWC limits forgetting to 3.8% versus 15.3% without EWC and 18.9% for LoRA; long-horizon degradation is 1.9% versus 27.6% for Full-Context, with 198 ms latency.
Methods
- Architecture: dual episodic/semantic memory, a gating MLP over recurrence, correction, task relevance, user feedback, and LM features; K-means plus LLM summarization for compression; retrieval scoring with cosine similarity, temporal decay, and NLI contradiction detection; EWC with utility-weighted experience replay.
- Data and baselines: PersonaChat-Long (50 sessions/user, 10,234 turns, 87 users), MSC (5 sessions × 1,800 dialogues), ConvAI2-Continual (8,500 turns), and SyntheticPersonal (12,000 turns, 200 users); compared against Vanilla LLM, Full-Context, RAG-Dense, RAG-BM25, LoRA-FT, and Memorizing Transformer using LLaMA-2-7B, BERT-base encoder, and RoBERTa-large-MNLI.
- Analysis: metrics include personalization accuracy, factual consistency, BLEU-4, ROUGE-L, memory efficiency, forgetting, and latency; theoretical bounds cover memory growth, consistency, rate-distortion, and convergence; experiments run on 4× NVIDIA A100 GPUs.
Limitations
- Stated: compression irreversibility permanently loses fine-grained details; cold-start performance is suboptimal in sessions 0–5; adversarial users could inject contradictions; long-term storage raises privacy/data-retention concerns; the 72 ms retrieval overhead may be prohibitive for latency-critical applications; and NLI-based consistency detection can fail on nuanced or domain-specific contradictions.
- Evident from the text: the main results table is fragmented across sections, with some baseline and MEMLLM rows appearing only in later table fragments, making direct comparison harder from the presented tables alone.
- Evident from the text: some reported numbers vary slightly across sections (e.g., 0.89 vs. 0.893 gating F1; 72.7% vs. 73% storage reduction; 68% vs. 77.6% contradiction reduction versus RAG-Dense), likely due to rounding or inconsistent reporting.
Model-written summary. Check it against the paper before relying on it.
Integrity signals
Registry
Cites
Statistics
Verification
Claims the paper makes
10 Valid · 4 Not valid · 3 Not checked · 98 Own results and interpretations
Some cited sources support this claim and some don’t — read each source’s verdict below.
In the paper
“MemN2N [20], NTM [21], and DNC [22] introduced differentiable external memory for neural networks. However, these focus on algorithmic tasks rather than long-term personalization.”
What the sources say
- Disputedcites [20] [20] S. Sukhbaatar, A. Szlam, J. Weston, and R. Fergus, 'End-to-end memory networks,' in NeurIPS , 2015, pp. 2440-2448. - sourceExperimental — accuracy not yet measured
The Abstract (p. 1) and Conclusions (p. 8) describe MemN2N's applications as synthetic question answering and language modeling. Related Work (p. 3) explicitly distinguishes its textual reasoning tasks from the abstract sorting and recall tasks tackled by the NTM. The paper supports MemN2N's use of external memory, but contradicts the characterization of MemN2N as focused on algorithmic tasks; it does not assess long-term personalization. (LLM, conf=0.86)
Checked by a reasoning model — may be wrong.
- Supportedcites [21] [21] A. Graves, G. Wayne, and I. Danihelka, 'Neural turing machines,' arXiv preprint arXiv:1410.5401 , 2014. - source
For the NTM citation, the Introduction (pp. 1–2) says the architecture was designed to simplify algorithmic tasks and likens its memory to short-term working memory. Section 3 (p. 5) describes its differentiable external memory. This supports the contrast with long-term personalization for NTM; the paper is not evidence about MemN2N or DNC. (LLM, conf=0.91)
- Not checked — no full text of the sourcecites [22] [22] A. Graves et al. , 'Hybrid computing using a neural network with dynamic external memory,' Nature , vol. 538, no. 7626, pp. 471-476, 2016. - source
plain- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: unclear — It is unclear whether the paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Disputed
The cited sources support only part of this claim — read each source’s verdict below.
In the paper
“Our continual learning framework with elastic weight consolidation maintains stable performance with only 3.8% degradation over extended interactions, conclusively solving the catastrophic forgetting problem for personalized LLMs.”
What the sources say
- Partly supportedcites [9] [9] J. Kirkpatrick et al. , 'Overcoming catastrophic forgetting in neural networks,' Proceedings of the National Academy of Sciences , vol. 114, no. 13, pp. 3521-3526, 2017. - sourceExperimental — accuracy not yet measured
The Introduction (p. 2) and Discussion (p. 6) describe elastic weight consolidation as a way to protect previously learned tasks during continual learning. The paper tests it on digit classification and Atari games, not MEMLLM or personalized LLMs; it does not report 3.8% degradation over extended interactions or establish a conclusive solution for that setting. (LLM, conf=0.90)
boostedstates a cause- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Partly supported
A result the paper reports from its own data — it cites no other source for it, so there is none to check. The quote shows it is in the paper.
In the paper
“Consistency plummets to 79.4% (14.3% absolute drop)”
plainstates a cause- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
A result the paper reports from its own data — it cites no other source for it, so there is none to check. The quote shows it is in the paper.
In the paper
“Extensive experiments on multi-session dialogue datasets spanning 10,000+ turns demonstrate that MEMLLM achieves 94.3% personalization accuracy while reducing memory footprint by 73% and factual contradictions by 68% compared to retrievalaugmented baselines.”
boosted- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
A result the paper reports from its own data — it cites no other source for it, so there is none to check. The quote shows it is in the paper.
In the paper
“Consistency : 93.7% consistency score”
plain- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
76 more claims
Create a free account to see all 115 claims and to verify the 1 reference still open.
- Probably not a substantial claim. 10
- It may state two separate claims as one. 27
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.
Rated major: it would touch a result — a value a result rests on, reported and recomputed p-values on opposite sides of the significance level, two passages that conflict with no sign which is right, or a cited work retracted before this paper was published.
In the paper
“Consistency : 93.7% consistency score, reducing contradictions by 68% compared to RAG-Dense (from 28.2% to 6.3% contradiction rate).”
Reported 68 · recomputed 77.659574
Rated major: it would touch a result — a value a result rests on, reported and recomputed p-values on opposite sides of the significance level, two passages that conflict with no sign which is right, or a cited work retracted before this paper was published.
In the paper
§ D. Long-Horizon Analysis
“While baselines exhibit severe performance degradation Full Context drops 27.6%, RAG-Dense 24.3%, Memorizing Transformer 15.8%, our method degrades only 1.9%, maintaining 93.2% accuracy.”
§ (b) Factual consistency over time E t S t
“Fig. 2: Comprehensive performance analysis. (a) MEMLLM maintains 93.2% accuracy at 10K turns vs. 56.8% for Full-Context (1.8% vs. 21.6% degradation).”
Rated minor: it would not touch a result — a value no result rests on, a p-value that stays on the same side of the significance level, or a conflict that is clerical, resolved elsewhere in the paper, or outside its own findings.
In the paper
§ I. Scalability Analysis
“Fig. 3: Storage scalability: MEMLLM plateaus at 240 MB/user while Full-Context grows linearly to 2.1 GB/user at 1000 users.”
§ I. Scalability Analysis
“Full-Context storage grows linearly ( O ( T ) ) reaching 2.1 GB per user at 1000 interactions.”
Rated minor: it would not touch a result — a value no result rests on, a p-value that stays on the same side of the significance level, or a conflict that is clerical, resolved elsewhere in the paper, or outside its own findings.
In the paper
“Catastrophic Forgetting : Minimal 3.8% performance degradation on old tasks, 80% better than LoRA (18.9%) and 75% better than Memorizing Transformer (11.4%).”
Reported 75 · recomputed 66.666667
Citations
Cites
The publisher hasn’t deposited this paper’s reference list.
Research works printed but not linked: 3
- [1][1] A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, Ł. Kaiser, and I. Polosukhin, 'Attention is all you need,' in Advances in Neural Information Processing Systems , 2017, pp. 5998-6008. -
- [17][17] E. J. Hu et al. , 'LoRA: Low-rank adaptation of large language models,' in ICLR , 2022. -
- [27][27] N. Mrkˇ si´ c et al. , 'Neural belief tracker: Data-driven dialogue state tracking,' in ACL , 2017, pp. 1777-1788. -
Metadata
- Authors
- Kummara, H.
- Affiliations
- University of Bath (GB)
- Corresponding author
- Kummara, Harshavardhan
- Venue
- Cambridge Open Engage (preprint)
- Publisher
- Cambridge University Press (CUP)
- Published
- 2026-02-01
- OpenAlex
- W7127051997
- Cited by
- 0
- References
- 0
- Type
- posted content
- Language
- en
- Open access
- green
- doi.org
- Funders
- University of Cambridge
- Topics
- Physical Sciences › Computer Science › Artificial Intelligence › Topic Modeling
- Physical Sciences › Computer Science › Artificial Intelligence › Speech and dialogue systems
- Physical Sciences › Computer Science › Computer Vision and Pattern Recognition › Multimodal Machine Learning Applications
- Keywords
- long-term memory, large language models, episodic memory, semantic memory, continual learning, catastrophic forgetting, dialogue coherence, retrieval-augmented generation, personalization, factual consistency
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.
The claim verdicts this report counts are experimental: their accuracy is not yet measured.
A language model drew the claims from the paper on 2026-10-01, under version 3 of the claim definition (model deepseek-v4.1-flash, prompt extract-v3).
A second model labelled each claim on 2026-10-01 (model jev-1.13-20260917, definition revision gold-guideline-v0.2).
Cited sources are checked with a language model’s help (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
How to cite
Cite CoChat for the checks on this page, and the paper for its own findings.
- This page
- CoChat. (2026). Citation check: From Goldfish to Elephant: Long-Term Memory for Large Language Models via Selective Storage and Hierarchical Compression. https://app.cochat.ai/papers/10.33774/coe-2026-xpkvb
- The paper
- Kummara, H. (2026). From Goldfish to Elephant: Long-Term Memory for Large Language Models via Selective Storage and Hierarchical Compression. Cambridge Open Engage (preprint). https://doi.org/10.33774/coe-2026-xpkvb
Data sources
Bibliographic data: Semantic Scholar, OpenAlex, Crossref.
Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).
Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.