MeMo: Cross-domain Memory & Modular Integration

• MeMo is a cross-domain research label denoting diverse methods of memory management, modular control, and component integration across various fields.
• It encompasses astronomical microlensing programs, memory-augmented neural networks, modular robotic controllers, and multimodal datasets in dialogue and biomedical imaging.
• The label bridges experimental systems and theoretical frameworks, yielding significant gains in efficiency, robustness, and overall performance.

MeMo, also written MEMO or Memo, is a recurrent acronym and project name in the arXiv literature rather than a single canonical method. It has been used for a combined microlensing program toward the Large Magellanic Cloud, single-sample test-time adaptation in vision, explicit associative-memory LLMs, long-range conversational memory systems, modular robot controllers, multimodal corpora, and systems techniques for long-context training and hybrid memory management (Mirhosseini et al., 2017, Zhang et al., 2021, Zanzotto et al., 18 Feb 2025, Lu et al., 2023, Tjandrasuwita et al., 2024, Tsfasman et al., 2024, Zhao et al., 2024). Across these usages, the label typically denotes either memory, modularity, or the integration of previously separate components.

1. Nomenclature and principal uses

The term has acquired multiple expansions across disciplines. In some papers it is an acronym; in others it functions as a project or corpus name. No single expansion is canonical across fields.

Usage	Expansion	Domain
MEMO project	MACHO, EROS, MOA, and OGLE considered together	Microlensing astronomy
MEMO	Marginal Entropy Minimization with One test point	Test-time adaptation
MEMO	A Deep Network for Flexible Combination of Episodic Memories	Memory-augmented neural reasoning
MeMo	Towards LLMs with Associative Memory Mechanisms	Language modeling
MeMo	Memory as a Model	Post-pretraining LLM knowledge integration
MEMO	Memory-Augmented MOdel context optimization	Multi-agent LLM games
MemoChat	self-composed memos for long-range conversation	Conversational LLMs
MeMo	Meaningful, Modular controllers via noise injection	Robot control transfer
MEMO	Memory Enhanced Manipulation	Neuro-symbolic robotics
MeMo corpus	Memory Modelling	Multiparty conversation dataset
MEMO dataset	EMA–OCTA retinal registration dataset	Biomedical imaging
MEMO	Fine-grained Tensor Management For Ultra-long Context LLM Training	LLM systems

This diversity has made MeMo a cross-domain research label. In astronomy it denotes a decades-long observational synthesis; in machine learning it often denotes explicit memory or adaptation mechanisms; in robotics it marks modular or retrieval-augmented control; and in systems work it names resource-management frameworks (Moniez et al., 2018, Banino et al., 2020, Quek et al., 14 May 2026, Christie et al., 4 Mar 2026, Liu et al., 2017).

2. Astronomical usage: the MEMO microlensing project

In astronomy, MEMO denotes the proposal to combine the historical microlensing surveys MACHO, EROS, MOA, and OGLE into a single long-baseline search toward the Large Magellanic Cloud for very long-duration microlensing events caused by heavy compact objects in the Galactic halo (Mirhosseini et al., 2017). The scientific target is a population of intermediate-mass or very massive black holes, including candidates around $10$–$10^2\,M_\odot$ and above, for which event durations extend to years rather than months.

The physical motivation is the standard microlensing scaling $t_E = R_E / v_T$, with $t_E \propto M^{1/2}$. The MEMO papers therefore emphasize that a $100\,M_\odot$ halo black hole produces events of order years, precisely where individual surveys of roughly 5–8 years lose efficiency (Mirhosseini et al., 2017). By combining the databases, the effective coverage becomes much longer: the overlap analysis identifies populations observed “almost continuously” for about 16, 21, 25, and 27 years, although the event-yield calculations are dominated by the 25-, 21-, and 16-year CCD-era combinations (Mirhosseini et al., 2017).

The key quantitative claim is that, under a standard spherical Galactic halo entirely composed of $100\,M_\odot$ objects, a MEMO-style combined analysis would expect about 15 events, compared with about 1 for EROS-2 alone and about 5 for OGLE-III alone (Moniez et al., 2018). The abstract of the earlier feasibility study summarizes the gain as “about ten more events caused by $100\,M_\odot$ black holes” relative to individual-survey searches if such objects make a major halo contribution (Mirhosseini et al., 2017).

The project is intentionally framed as feasible but technically demanding. The surveys differ in passbands, cadence, blending, source extraction, and calibration. The authors explicitly identify the need for color equations between passbands, accurate instrumental zero-point calibration, survey-specific treatment of photometric uncertainties and blending, and simulations of combined light curves before a final global detection efficiency can be established (Mirhosseini et al., 2017). MEMO in this sense is both an astrophysical search program and a cross-survey data-integration problem.

3. Memory and adaptation in neural architectures

Several MeMo/MEMO papers in machine learning use the label for methods that alter inference-time behavior or explicit memory structure. A prominent example is “MEMO: Test Time Robustness via Adaptation and Augmentation” (Zhang et al., 2021). That method addresses the strict one-test-point setting: for a single input $x$, it samples augmentations, averages the predictive distributions, minimizes the entropy of the marginal prediction over augmentations, updates all model parameters with one gradient step, and then predicts on the original input. The paper reports accuracy gains of 1–8\% over standard model evaluation and state-of-the-art single-test-point results on ImageNet-C, ImageNet-R, and, among ResNet-50 models, ImageNet-A (Zhang et al., 2021). Its defining objective is not ordinary conditional entropy minimization but entropy minimization of the averaged prediction across augmentations.

A different neural use appears in “MEMO: A Deep Network for Flexible Combination of Episodic Memories” (Banino et al., 2020). There, MEMO is an external-memory architecture designed for long-distance associative reasoning. The two distinguishing ideas are a separation between facts stored in memory and the items composing those facts, and an adaptive retrieval mechanism that permits a variable number of memory hops before answering. On the paper’s paired associative inference task, MEMO is the only compared architecture that reliably solves longer indirect chains; on the joint 10k bAbI benchmark it reports 0.21 mean error with 20/20 solved (Banino et al., 2020).

The explicit-memory direction is pushed further in “MeMo: Towards LLMs with Associative Memory Mechanisms” (Zanzotto et al., 18 Feb 2025). That work replaces standard training-centric memorization with direct storage of token-sequence associations in layered correlation matrix memories, using the classical associative form

$C = \sum_{i=1}^n \vec{k}_i \vec{v}_i^T.$

The architecture is designed to memorize text directly, supports explicit forgetting by subtraction, and experimentally studies both one-layer and multi-layer configurations (Zanzotto et al., 18 Feb 2025). In contrast to transformer-based implicit storage, memory in this formulation is inspectable and editable by design.

Taken together, these papers use MeMo/MEMO to denote architectures in which memory is no longer a loose metaphor. It is either an optimization target at inference time, an explicit external retrieval structure, or a directly writable associative store.

4. Language-model systems: conversation, knowledge integration, and multi-agent play

A distinct branch of the MeMo literature concerns LLM interaction protocols rather than neural architecture alone. “MemoChat: Tuning LLMs to Use Memos for Consistent Long-Range Open-Domain Conversation” (Lu et al., 2023) trains chat models to write and use structured topic-level memos in an iterative memorization–retrieval–response loop. The memo format is JSON-like, with fields such as "topic", "summary", and dialogue spans. On the paper’s expert-annotated MT-Bench+ benchmark, MemoChat-ChatGPT achieves an average GPT-4-judged score of 70.76, compared with 51.89 for ChatGPT-2k, 54.52 for MPC-ChatGPT, and 42.44 for MemoryBank-ChatGPT (Lu et al., 2023). The central claim is that topical self-authored memos are a better intermediate representation for long-range open-domain dialogue than raw context-window replay.

“MeMo: Memory as a Model” (Quek et al., 14 May 2026) generalizes memory to a modular post-pretraining knowledge-integration framework. A frozen Executive LLM remains unchanged, while a separate smaller Memory LM is fine-tuned on a synthesized reflection QA corpus derived from the target documents. At inference time, the Executive decomposes the user query, queries Memory for compact natural-language answers, and synthesizes the final response. On BrowseComp-Plus, NarrativeQA, and MuSiQue, MeMo reports strong results relative to retrieval baselines; with Qwen2.5-32B-Instruct as Executive, the reported accuracies are 54.22, 26.85, and 48.30, respectively, and with Gemini-3-Flash they rise to 66.67, 53.58, and 60.20 (Quek et al., 14 May 2026). The paper’s main systems claim is that retrieval cost is independent of corpus size at inference time because the deployed memory is the trained Memory model itself rather than a raw-corpus retriever.

Multi-agent evaluation produces another usage in “MEMO: Memory-Augmented MOdel context optimization for Robust Multi-Turn Multi-Agent LLM Games” (Xie et al., 9 Mar 2026). There, the optimized object is inference-time context $c = (q, M)$, where $10^2\,M_\odot$0 is the prompt and $10^2\,M_\odot$1 is a memory payload drawn from a persistent bank of structured self-play insights. Selection uses a lower-confidence TrueSkill score,

$10^2\,M_\odot$2

to favor reliable rather than merely lucky prompt variants. Across five text-based games and 2,000 self-play games per task, MEMO raises mean win rate from 25.1\% to 49.5\% for GPT-4o-mini and from 20.9\% to 44.3\% for Qwen-2.5-7B-Instruct, while also reducing run-to-run variance (Xie et al., 9 Mar 2026).

These LLM-oriented MeMo systems share a common pattern: memory is externalized, structured, and used to scaffold inference, but the base model is often left unchanged.

5. Multimodal conversation, speech, video, and biomedical imaging

The name also appears in multimodal corpora and perceptual-generation systems. “Introducing MeMo: A Multimodal Dataset for Memory Modelling in Multiparty Conversations” (Tsfasman et al., 2024) presents the first conversational dataset annotated with participants’ own memory retention reports. The corpus contains 31 hours of small-group online discussions about Covid-19, repeated 3 times over 2 weeks, with 53 participants across 15 groups. In the curated memory data there are 602 remembered moments, with an average of 3.9 moments per person, average description length of 32 words, and mean annotated duration of 141 seconds (Tsfasman et al., 2024). Its importance lies less in algorithmic novelty than in the explicit linking of remembered conversational events to the original temporal spans.

In speech separation, “MeMo: Attentional Momentum for Real-time Audio-visual Speaker Extraction under Impaired Visual Conditions” (Li et al., 21 Jul 2025) proposes two adaptive memory banks—a speaker bank and a contextual bank—that store information from previously extracted speech so the system can continue tracking a target speaker even when current visual cues are missing or degraded. The framework is designed for online streaming with default inference parameters $10^2\,M_\odot$3 s, $10^2\,M_\odot$4 s, and $10^2\,M_\odot$5 s. On the TDSE backbone under impaired visuals in online mode, SI-SNR improves from 8.13 dB to 10.34 dB with the contextual bank; similar gains of at least 2 dB are reported on USEV and BSRNN (Li et al., 21 Jul 2025).

In generative video, “MEMO: Memory-Guided Diffusion for Expressive Talking Video Generation” (Zheng et al., 2024) combines a memory-guided temporal module with an emotion-aware audio module in a latent diffusion architecture for audio-driven portrait animation. The memory states store long-range temporal information through linear-attention summaries, while emotion is inferred from audio and injected through emotion-adaptive conditioning. On the paper’s VoxCeleb2 test set, MEMO reports FVD 254.3, FID 31.7, and Sync-D 7.4; on a collected out-of-distribution dataset, it reports FVD 161.1, FID 24.9, and Sync-D 9.2 (Zheng et al., 2024).

Biomedical imaging supplies yet another meaning. “MEMO: Dataset and Methods for Robust Multimodal Retinal Image Registration with Large or Small Vessel Density Differences” (Wang et al., 2023) introduces the first public paired EMA–OCTA retinal registration dataset, containing 30 image pairs from four eyes of two healthy rhesus macaques, together with a density-aware registration method, VDD-Reg, and the Masked Soft Dice metric. The method is explicitly engineered for large inter-modality vessel-density differences and maintains its success rate with as few as three annotated EMA vessel masks (Wang et al., 2023).

6. Robotics, modular control, and quantum circuit design

In robotics, MeMo frequently denotes modularity plus reuse. “MeMo: Meaningful, Modular Controllers via Noise Injection” (Tjandrasuwita et al., 2024) begins from a single trained monolithic robot/controller pair and distills it into a hierarchical controller with a global boss and shared local worker modules tied to physical subassemblies. The core regularizer injects Gaussian noise into the boss-to-worker latent communication path during imitation learning so that modules become insensitive to many boss-signal directions and therefore assume more of the low-level coordination burden themselves. On morphology transfer tasks such as 6→12 leg centipede and 6→10 leg hybrid, the paper reports that MeMo is about $10^2\,M_\odot$6 more sample efficient than the best baseline and also reaches better final performance (Tjandrasuwita et al., 2024).

“From Local Corrections to Generalized Skills: Improving Neuro-Symbolic Policies with MEMO” (Christie et al., 4 Mar 2026) uses the acronym for Memory Enhanced Manipulation. Here the memory artifact is a retrieval-augmented skillbook built from human feedback and successful executions. Free-form corrections are paraphrased, clustered, and fused with successful code templates into generalized reusable guidance. In simulation on held-out tasks, MEMO reaches 78\% zero-shot success, compared with 40\% for MEMO-S/TrajGen and 28\% for DROC-V; in the real world it reports 88\% overall success with 1.52 feedback items per task on average (Christie et al., 4 Mar 2026). The method is explicitly aimed at expanding the robot’s effective skill repertoire without retraining a monolithic policy.

A more specialized optimization usage appears in “MEMO-QCD: Quantum Density Estimation through Memetic Optimisation for Quantum Circuit Design” (Ardila-García et al., 2024). That work applies a memetic algorithm—genetic architecture search plus local gradient-based parameter tuning—to design shallow quantum feature-map circuits that approximate Gaussian kernel density estimation. The memory aspect here is nominally absent; MeMo refers to memetic optimisation, not episodic or retrieval memory. The paper’s emphasis is on feasibility for near-term hardware through shallow circuit design (Ardila-García et al., 2024).

Across these works, MeMo marks either modular factorization of control or hybrid optimization of reusable structures, rather than a single robotics formalism.

7. Systems software and engineering-process usages

Systems papers use the label for memory management in the literal computer-systems sense. “MEMO: Fine-grained Tensor Management For Ultra-long Context LLM Training” (Zhao et al., 2024) is a training framework for ultra-long-context transformers that offloads memory-consuming activations to CPU memory, recomputes only selected token slices, and uses a bi-level mixed-integer program to optimize memory reuse and reduce fragmentation. The abstract reports average MFU improvements of 1.97× over Megatron-LM and 1.80× over DeepSpeed, and states that MEMO trains a 7B model with 1 million sequence length on 8 A800 GPUs at 52.30\% MFU (Zhao et al., 2024). The paper’s systems significance lies in treating activation management and fragmentation as first-class optimization problems rather than relying primarily on recomputation or communication-heavy parallelism.

The older OS-level framework “Memos: Revisiting Hybrid Memory Management in Modern Operating System” (Liu et al., 2017) uses a near-homographic spelling for an entirely different purpose: scheduling data placement across cache, channels, DRAM, and NVM in a hybrid memory hierarchy. The reported gains are substantial: average throughput improvement of 19.1\%, QoS improvement of 23.6\%, NVM-side memory latency reduction of 3–83.3\%, energy reduction of 25.1–99\%, and average NVM lifetime improvement of 40× (Liu et al., 2017). Here the central concern is not model memory but operating-system placement and migration policy.

At the process-methodology level, “Mesh Model (MeMo): A Systematic Approach to Agile System Engineering” (Mishra, 2017) defines MeMo as a hybrid engineering process combining spiral-style iteration with Design Blocks and Feedback Collection Blocks. It is a conceptual white paper rather than an algorithmic system, but it preserves the same nominal theme of combining structured design with reusable feedback knowledge (Mishra, 2017).

In aggregate, these system-oriented usages show that MeMo/MEMO can refer to runtime resource orchestration, operating-system memory hierarchies, or even engineering-process structure. The commonality is therefore lexical rather than methodological.