U-Mem: Memory-Centric Computational Models

• U-Mem is a family of memory-centric constructs that treats memory as an active computational substrate, integrating hardware, formal computation, and adaptive agent design.
• It encompasses varied implementations including filamentary mem-resistors, unimodal membership inference detectors, and unified memory agents optimized for long-context reasoning.
• U-Mem’s applications span autonomous memory agents, privacy-preserving inference, and compact model architectures, delivering performance improvements and operational efficiencies.

U-Mem denotes a family of memory-centered constructs rather than a single canonical method. In the cited literature, the label appears explicitly in autonomous memory agents, is proposed as a natural shorthand for unimodal membership inference and several unified-memory agent designs, and also has a hardware and formal-computation lineage in memristive systems and memcomputing (Wu et al., 25 Feb 2026, Cheng et al., 15 Mar 2026, Zhang et al., 13 Feb 2026, Ye et al., 11 Feb 2026, Mouttet, 2011, Pei et al., 2017). Taken together, these works indicate that U-Mem is best understood as an umbrella for approaches in which memory is an active computational substrate, not merely a passive repository.

1. Terminological scope

The most important preliminary fact is terminological: U-Mem is not standardized across the literature. Different papers use the name directly, use closely related names, or explicitly note that U-Mem is a natural shorthand for the mechanism under study.

Research line	Canonical term in the paper	Defining feature
Memristive hardware	Filamentary mem-resistor; UMM (Mouttet, 2011, Pei et al., 2017)	Dynamic tunneling and Schottky barriers; memory both stores and processes information
Multimodal privacy auditing	UMID (Cheng et al., 15 Mar 2026)	Text-only membership inference via cross-modal latent inversion
Long-context reasoning	UMA; UMEM (Zhang et al., 13 Feb 2026, Ye et al., 11 Feb 2026)	Unified memory operations, dual memory banks, joint extraction and management
Interactive agents	UI-Mem; O-Mem (Xiao et al., 5 Feb 2026, Wang et al., 17 Nov 2025)	Self-evolving experience memory; active user profiling
Compact or sparse model memory	$\delta$-mem; UltraMem; U-Hop (Lei et al., 12 May 2026, Huang et al., 2024, Wu et al., 2024)	Fixed-size online state, ultra-sparse memory layers, uniform associative retrieval
Autonomous memory agents	U-Mem (Wu et al., 25 Feb 2026)	Cost-aware knowledge acquisition and semantic-aware Thompson sampling

A common misconception is to read U-Mem as if it named one specific architecture. The literature instead uses the label for several distinct but structurally related ideas: memristive state dynamics, external memory for agents, test-time memory adaptation, and privacy-sensitive inference over memory effects.

2. Hardware and formal-computation lineage

In the hardware lineage most directly associated with a memristive interpretation of U-Mem, filamentary mem-resistors are modeled as a nonlinear dynamic system whose state is jointly determined by a dynamic tunneling barrier and a dynamic Schottky barrier. The two principal state variables are the filament-electrode gap $x_f(t)$ and the depletion width $x_d(t)$. The total current density is

$J(t)=\frac{[J_s(t)+J_c(t)](A-A_f)+J_T(t)A_f}{A},$

so the device combines Schottky conduction, tunneling conduction, and capacitive displacement current in a single state-space description. Under sinusoidal drive, the gap dynamics reduce near equilibrium to a driven, damped harmonic oscillator, and the model explicitly predicts a zero-crossing hysteresis curve that disappears when the input frequency moves sufficiently far from resonance. This framework was presented as directly relevant to any U-Mem concept built on filamentary RRAM/ReRAM (Mouttet, 2011).

At the abstract computational level, Universal Memcomputing Machines formalize the idea that memory itself performs computation. In the original formulation, a UMM is an eight-tuple

$(M,\Delta,P,S,\Sigma,p_0,s_0,F),$

while the later set-theoretic abstraction represents a machine as $S \times \Phi'_S$, where $S$ is the internal state set and $\Phi'_S$ is the realizable transition set. For UMMs, the defining idealization is a continuum-sized state space with $|M|=\beth_1$ together with the full transition set $\Phi_M$. On that basis, the model is shown to be Turing-complete, liquid-complete, and quantum-complete. The formalism also emphasizes intrinsic parallelism, information overhead, and functional polymorphism, all of which place memory at the center of the computational model rather than at its periphery (Pei et al., 2017).

A related associative-memory formulation appears in U-Hop, which can be read as a uniform-memory construction. U-Hop introduces a learnable feature map $x_f(t)$0 and a separation loss

$x_f(t)$1

followed by standard Hopfield energy minimization in kernel space. The purpose is to make stored memories more uniformly separated, reduce metastable states, and enlarge effective associative-memory capacity. This suggests a formal analogue of U-Mem in which the key operation is not storage alone but controlled shaping of the attractor geometry of memory retrieval (Wu et al., 2024).

3. U-Mem as unimodal membership inference

A modern privacy-oriented usage treats U-Mem as unimodal membership inference for multimodal contrastive models. In this setting, the paper introduces UMID, the Unimodal Membership Inference Detector, and explicitly notes that “U-Mem” is a natural shorthand for the problem it studies: text-only auditing of whether a text-image or text-audio PII pair was present in the training data of CLIP- or CLAP-type models. The threat model is identity-level membership inference under a unimodal privacy constraint: the auditor may query only with text, has encoder and gradient access, does not train shadow models, and aims for low per-query cost (Cheng et al., 15 Mar 2026).

The core mechanism is text-guided cross-modal latent inversion. For a text identity $x_f(t)$2, the text embedding is $x_f(t)$3. Starting from random noise $x_f(t)$4, the method performs gradient ascent on cosine similarity

$x_f(t)$5

producing $x_f(t)$6 optimized modality embeddings $x_f(t)$7. Two statistics are then extracted:

$x_f(t)$8

where $x_f(t)$9. High similarity and low variability indicate a stable identity-specific attractor in the latent space; low similarity and high variability indicate non-membership.

UMID estimates the non-member region using synthetic gibberish texts and applies an ensemble of unsupervised anomaly detectors—Isolation Forest, Local Outlier Factor, One-Class SVM, and Autoencoder—over the two-dimensional feature space $x_d(t)$0. In the reported implementation, $x_d(t)$1, $x_d(t)$2, $x_d(t)$3, and $x_d(t)$4 gibberish strings. With text-only auditing on CLIP ResNet-50 in the one-shot setting, UMID reports precision $x_d(t)$5, recall $x_d(t)$6, accuracy $x_d(t)$7, and time $x_d(t)$8 s/query; with one local real image kept off-model, accuracy rises to $x_d(t)$9. On LibriSpeech one-shot CLAP, text-only UMID reports accuracy $J(t)=\frac{[J_s(t)+J_c(t)](A-A_f)+J_T(t)A_f}{A},$0, and with one local audio sample it reports $J(t)=\frac{[J_s(t)+J_c(t)](A-A_f)+J_T(t)A_f}{A},$1. The paper also reports that differential privacy noise with $J(t)=\frac{[J_s(t)+J_c(t)](A-A_f)+J_T(t)A_f}{A},$2 reduces accuracy by roughly $J(t)=\frac{[J_s(t)+J_c(t)](A-A_f)+J_T(t)A_f}{A},$3–$J(t)=\frac{[J_s(t)+J_c(t)](A-A_f)+J_T(t)A_f}{A},$4 percentage points, and that covert gibberish defenses reduce accuracy only slightly. A central limitation is that the method requires gradient access and is therefore gray-box rather than strict black-box.

4. Unified memory for long-context reasoning and self-evolving agents

In long-context reasoning, one important U-Mem interpretation is the Unified Memory Agent. UMA formulates streaming reasoning as an MDP with state $J(t)=\frac{[J_s(t)+J_c(t)](A-A_f)+J_T(t)A_f}{A},$5, where the memory is dual:

$J(t)=\frac{[J_s(t)+J_c(t)](A-A_f)+J_T(t)A_f}{A},$6

The core memory $J(t)=\frac{[J_s(t)+J_c(t)](A-A_f)+J_T(t)A_f}{A},$7 is a compact global summary, while $J(t)=\frac{[J_s(t)+J_c(t)](A-A_f)+J_T(t)A_f}{A},$8 is a structured key-value Memory Bank supporting explicit CRUD operations through Add, Update, Delete, Retrieve, List, and UpdateCore. Crucially, memory maintenance and question answering are handled by a single policy rather than separate subsystems. Training uses Task-Stratified Group Relative Policy Optimization, with separate advantages for memory trajectories and QA trajectories, so that early memory operations receive long-horizon credit from later question-answering performance. UMA introduces Ledger-QA as a benchmark for persistent state tracking and reports $J(t)=\frac{[J_s(t)+J_c(t)](A-A_f)+J_T(t)A_f}{A},$9 average accuracy across 13 datasets, versus $(M,\Delta,P,S,\Sigma,p_0,s_0,F),$0 for MemAlpha, $(M,\Delta,P,S,\Sigma,p_0,s_0,F),$1 for RAG, and $(M,\Delta,P,S,\Sigma,p_0,s_0,F),$2 for MemAgent. On Ledger-QA with 50 sessions, UMA reports $(M,\Delta,P,S,\Sigma,p_0,s_0,F),$3 accuracy, compared with $(M,\Delta,P,S,\Sigma,p_0,s_0,F),$4 for RAG and $(M,\Delta,P,S,\Sigma,p_0,s_0,F),$5 for MemAgent (Zhang et al., 13 Feb 2026).

A second line, UMEM, treats external memory as the trainable parameters of the agent while keeping the executor frozen. The memory bank is

$(M,\Delta,P,S,\Sigma,p_0,s_0,F),$6

and a Mem-Optimizer policy jointly outputs both the extracted memory content and the management operation: $\Phi_M$1 Its distinguishing feature is Semantic Neighborhood Modeling: for each query $(M,\Delta,P,S,\Sigma,p_0,s_0,F),$7, the framework constructs $(M,\Delta,P,S,\Sigma,p_0,s_0,F),$8 as the Top-$(M,\Delta,P,S,\Sigma,p_0,s_0,F),$9 nearest neighbors in embedding space, with $S \times \Phi'_S$0 in the main experiments, and evaluates a proposed memory update not on the source query alone but across the neighborhood. The reward is a neighborhood-level marginal utility that combines success gain and efficiency regularization, optimized by GRPO. The paper reports up to a $S \times \Phi'_S$1 improvement in multi-turn interactive tasks and states that UMEM maintains a monotonic growth curve during continuous evolution. Ablations further show that removing extraction optimization causes larger degradation than removing management optimization, which suggests that in unified-memory systems memory quality is often the primary bottleneck (Ye et al., 11 Feb 2026).

These two frameworks reject the view that memory is merely retrieval-time augmentation. Both instead treat memory as a learned control problem involving explicit update rules, structured storage, and task-conditioned reorganization.

5. Self-evolving external memory in interactive agents

In mobile GUI agents, UI-Mem defines U-Mem as a self-evolving hierarchical experience memory for online reinforcement learning. The memory is explicitly stratified into high-level workflows, mid-level subtask skills, and failure patterns, all stored as parameterized templates with placeholders. Its training-time contribution is Stratified Group Sampling, which mixes strong-guidance, weak-guidance, and no-guidance trajectories inside the same GRPO group to preserve outcome diversity while encouraging the policy to internalize guided behaviors. The system also uses a Self-Evolving Loop that abstracts new successful trajectories into workflows and skills, and failed trajectories into failure patterns. On AndroidWorld, UI-Mem-8B reports $S \times \Phi'_S$2 success without inference-time retrieval and $S \times \Phi'_S$3 with retrieval; on AndroidLab, UI-Mem-8B* reports Sub-SR $S \times \Phi'_S$4, RRR $S \times \Phi'_S$5, ROR $S \times \Phi'_S$6, and SR $S \times \Phi'_S$7. Component ablations show substantial drops when hierarchy, abstraction, self-evolution, or stratified sampling are removed (Xiao et al., 5 Feb 2026).

O-Mem addresses a different setting: personalized, long-horizon assistants. Its memory architecture separates persona memory $S \times \Phi'_S$8, working memory $S \times \Phi'_S$9, and episodic memory $S$0. Persona memory stores long-term attributes and factual events; working memory maps topics to interactions; episodic memory maps clue words to interactions,

$S$1

Updates are driven by active user profiling, in which each interaction is processed to extract topic, attribute, and event, followed by LLM-mediated Add/Ignore/Update decisions and nearest-neighbor clustering over persona attributes. Retrieval is parallel: topic-based lookup for working memory, inverse-document-frequency-style clue selection for episodic memory, and semantic retrieval over persona attributes and facts. O-Mem reports $S$2 on LoCoMo and $S$3 on PERSONAMEM, with average token cost $S$4K and delay $S$5 s, compared with direct RAG at $S$6K tokens and $S$7 s. Relative to LangMem on LoCoMo, O-Mem reports roughly a $S$8-point F1 improvement while reducing token consumption by $S$9 and latency by $\Phi'_S$0 (Wang et al., 17 Nov 2025).

Both UI-Mem and O-Mem are specialized rather than universal: the former is optimized for long-horizon GUI control under sparse rewards, while the latter is optimized for personalized dialogue and research assistance. Their commonality is architectural, namely explicit separation of memory types and continual memory evolution.

6. Autonomous memory agents and compact memory mechanisms

The most direct contemporary use of the name appears in autonomous memory agents. Here U-Mem is explicitly glossed as “Utility-driven Memory” and is defined by a Retrieve–Infer–Evolve loop over a frozen LLM and an external memory store. Its key novelties are a cost-aware knowledge-extraction cascade and semantic-aware Thompson sampling. For verifiable tasks, when an answer is wrong, the agent escalates from a stronger teacher LLM, to teacher plus tools, and only then to an expert; it then performs contrastive reflection between failed and correct trajectories to extract global procedural memories and local corrective memories. Retrieval treats each memory as a Gaussian utility posterior $\Phi'_S$1, samples utilities by Thompson sampling, and combines sampled utility with semantic similarity. The update signal is an advantage reward,

$\Phi'_S$2

so utility is tied to marginal contribution rather than task difficulty. U-Mem reports improvements of $\Phi'_S$3 points on HotpotQA for Qwen2.5-7B and $\Phi'_S$4 points on AIME25 for Gemini-2.5-flash, and an ablation shows that the full cascade attains $\Phi'_S$5 on HotpotQA with only $\Phi'_S$6 expert calls, versus $\Phi'_S$7 for always using the expert (Wu et al., 25 Feb 2026).

A very different but conceptually adjacent memory mechanism is $\Phi'_S$8-mem. Instead of an external store, it augments a frozen full-attention backbone with a compact online associative-memory state

$\Phi'_S$9

with default $|M|=\beth_1$0. The state is updated by a gated delta rule and read at every token to generate low-rank corrections to the attention computation. With only an $|M|=\beth_1$1 online memory state, $|M|=\beth_1$2-mem reports an average score $|M|=\beth_1$3 that of the frozen backbone and $|M|=\beth_1$4 that of the strongest non-$|M|=\beth_1$5-mem memory baseline; on memory-heavy tasks it reaches $|M|=\beth_1$6 on MemoryAgentBench and $|M|=\beth_1$7 on LoCoMo while largely preserving general capabilities. This suggests a compact recurrent interpretation of U-Mem in which memory is a tiny online state directly coupled to attention rather than a retrieved external corpus (Lei et al., 12 May 2026).

At the other end of the scale, UltraMem treats memory as an ultra-sparse replacement for Transformer FFN layers. It uses a large key-value memory layer with Tucker Decomposed Query–Key Retrieval, Implicit Value Expansion, and Multi-Core Scoring, and the largest model reported contains $|M|=\beth_1$8 million memory slots. The design goal is to decouple parameter count not only from FLOPs but also from inference-time memory access. Empirically, UltraMem-1.6B-$|M|=\beth_1$9 matches or exceeds a 6.5B dense model under a comparable computational budget, and at realistic batch sizes it is reported to be up to $\Phi_M$0 faster than MoE while keeping inference time close to a dense model (Huang et al., 2024).

Taken together, these strands show why U-Mem cannot be reduced to a single recipe. In one line of work it is an autonomous, cost-aware knowledge-acquisition system; in another it is a compact associative state or an ultra-sparse memory layer; in yet another it is a unified external memory controller for long-context agents. The shared thesis is narrower and more robust than the name itself: memory is treated as a first-class mechanism for computation, adaptation, control, and retrieval.