FluxMem: Adaptive Memory Systems

• FluxMem is a polysemous term denoting adaptive memory systems with domain-specific implementations in streaming video, LLM agents, and superconducting hardware.
• The training-free hierarchical memory for streaming video uses TAS and SDC modules to compress tokens, cutting latency and GPU memory usage significantly.
• In agentic systems, FluxMem adapts memory structures or evolves connectivity through probabilistic gating and LLM-guided refinement to boost performance.

Searching arXiv for papers using the term “FluxMem” and closely related usages to ground the article in current literature. {"query":"all:FluxMem","max_results":10,"sort_by":"submittedDate","sort_order":"descending"} Relevant arXiv results using or discussing “FluxMem” include:

• "Rethinking Memory as Continuously Evolving Connectivity" (Fang et al., 27 May 2026)
• "FluxMem: Adaptive Hierarchical Memory for Streaming Video Understanding" (Xie et al., 2 Mar 2026)
• "Choosing How to Remember: Adaptive Memory Structures for LLM Agents" (Lu et al., 15 Feb 2026)

I will use these directly, and only use additional papers where the supplied data explicitly links them to the term or to flux-controlled memory concepts. FluxMem is a recurrent research label rather than a single canonical architecture. In current arXiv usage, it most prominently denotes three distinct classes of memory systems: a training-free, plug-and-play hierarchical memory for streaming video understanding, a unified framework for adaptive memory organization in LLM agents, and a connectivity-evolving memory framework for agentic systems. In adjacent superconducting-memory discourse, the label is also used more loosely for flux-controlled or flux-activated memory mechanisms rather than for a standardized software framework (Xie et al., 2 Mar 2026, Lu et al., 15 Feb 2026, Fang et al., 27 May 2026, Guarcello et al., 2017, Valadares et al., 20 Feb 2026).

1. Terminological scope

The main scholarly use of the name is concentrated in 2026 work on memory systems for multimodal or agentic AI, but the same label also appears in technical exposition around superconducting memory concepts. This makes “FluxMem” a polysemous term whose meaning depends on domain.

Usage of “FluxMem”	Domain	Defining characteristics
"FluxMem: Adaptive Hierarchical Memory for Streaming Video Understanding" (Xie et al., 2 Mar 2026)	Video–LLM systems	training-free, plug-and-play hierarchical memory; TAS and SDC; self-adaptive token compression via Otsu’s method
"Choosing How to Remember: Adaptive Memory Structures for LLM Agents" (Lu et al., 15 Feb 2026)	LLM agents	multiple complementary memory structures; three-level hierarchy; selector trained from downstream rewards; BMM-based probabilistic gate
"Rethinking Memory as Continuously Evolving Connectivity" (Fang et al., 27 May 2026)	Agentic systems	heterogeneous graph memory; three-stage evolution; feedback-driven refinement; PEMS-guided consolidation
Flux-controlled memory concepts (Guarcello et al., 2017, Valadares et al., 20 Feb 2026)	Superconducting hardware	temperature-based superconducting memory element; flux-activated resonant control of a bosonic quantum memory

A common misconception is that FluxMem refers to one standardized memory stack. The literature instead uses the name for technically different proposals that share an emphasis on adaptive retention, controllable compression, or explicitly managed memory evolution.

2. Adaptive hierarchical memory for streaming video understanding

In "FluxMem: Adaptive Hierarchical Memory for Streaming Video Understanding" (Xie et al., 2 Mar 2026), FluxMem is a training-free framework for efficient streaming video understanding. It is described as a training-free, plug-and-play hierarchical memory for video–LLM systems, designed to work with existing MLLMs—primarily Qwen2.5‑VL‑7B—without retraining. The central problem is that modern video MLLMs encode each frame into hundreds of visual tokens, and under streaming conditions the resulting token stream quickly blows up latency and GPU memory.

The system maintains three levels of memory over time: short-term, mid-term, and long-term. Recent frames are kept dense in short-term memory, while older frames undergo progressively stronger compression. When a frame $F_t$ arrives, a vision encoder produces tokens

$V_t \in \mathbb{R}^{HW \times D},$

which are written into short-term memory $\mathcal{M}^s$, mid-term memory $\mathcal{M}^m$, and long-term memory $\mathcal{M}^l$ according to capacity-driven promotion. Evicted short-term tokens are compressed and moved to mid-term; evicted mid-term tokens are further compressed and moved to long-term.

The architecture uses two modules. Temporal Adjacency Selection (TAS) removes redundant visual tokens across adjacent frames. Spatial Domain Consolidation (SDC) merges repetitive spatial regions within a frame into compact anchor representations. The temporal novelty criterion in TAS uses cosine distance

$d(x,y)=1-\cos(x,y),$

with backward and forward novelty scores defined over a $3\times 3$ neighborhood in adjacent frames. A token is retained when

$$\text{keep}(t,h,w)\iff (s_{t,h,w}^{-}>\Theta_t^{-}) \lor (s_{t,h,w}^{+}>\Theta_t^{+}).$$

The thresholds $\Theta_t^{-}$, $\Theta_t^{+}$, and the spatial threshold used by SDC are not manually tuned; they are obtained via Otsu’s method from per-frame similarity distributions. This self-adaptive token compression mechanism is a defining feature of the method.

SDC operates after TAS on retained tokens, constructs a sparse 8-connected graph over neighboring tokens in the same frame, extracts connected components with Union-Find, and replaces each component with a mean anchor token

$V_t \in \mathbb{R}^{HW \times D},$0

This yields a hierarchy in which short-term memory is dense, mid-term memory is motion-focused, and long-term memory is highly compressed and spatially consolidated.

The reported empirical behavior is explicitly strong in both online and offline settings. Under real-time settings, FluxMem reaches 76.4 on StreamingBench and 67.2 on OVO-Bench, while reducing latency by 69.9% and peak GPU memory by 34.5% on OVO-Bench (Xie et al., 2 Mar 2026). It also maintains offline performance, achieving 73.1 on MLVU while using 65% fewer visual tokens. On OVO-Bench, latency is reduced from 2701 ms to 812 ms and peak GPU memory from 35.8 GB to 23.5 GB. On MLVU, latency drops from 3614 ms to 2014 ms and peak GPU memory from 41.3 GB to 28.4 GB. The reported per-frame online overhead is approximately 4.1 ms, broken down as TAS: 1.3 ms, SDC: 2.4 ms, and Other logic: 0.4 ms.

The method is also explicitly query-agnostic. This suggests a design philosophy in which long-horizon visual retention is optimized at inference time through scene statistics rather than through task-specific learned policies. A plausible implication is that FluxMem, in this sense, is best understood as a memory-management layer inserted between the vision encoder and the LLM rather than as a new multimodal backbone.

3. Adaptive memory structures for LLM agents

In "Choosing How to Remember: Adaptive Memory Structures for LLM Agents" (Lu et al., 15 Feb 2026), FluxMem denotes a unified framework for LLM agents that adapts not only what is stored, but also how memory is organized. The paper identifies two limitations of existing agent memory systems: reliance on a one-size-fits-all memory structure and failure to model memory-structure selection as a context-adaptive decision.

This FluxMem organizes memory into a three-level hierarchy: $V_t \in \mathbb{R}^{HW \times D},$1 The levels are Short-Term Interaction Memory (STIM), Mid-Term Episodic Memory (MTEM), and Long-Term Semantic Memory (LTSM). STIM is a recency-weighted working buffer with capacity $V_t \in \mathbb{R}^{HW \times D},$2 pages and LRU-style eviction. MTEM stores episodic sessions, and each session uses exactly one of three structures: linear, graph, or hierarchical. LTSM stores abstracted, stable knowledge such as user profile, preferences, and long-term habits.

The system explicitly learns a structure selector. At each turn $V_t \in \mathbb{R}^{HW \times D},$3, it constructs a feature vector $V_t \in \mathbb{R}^{HW \times D},$4 with $V_t \in \mathbb{R}^{HW \times D},$5 interaction-level features, including page_count, avg_page_length, entity_density, relation_indicators, topic_diversity, topic_transitions, is_qna_pattern, is_decision_tree, is_entity_centric, time_span, temporal_density, and semantic_complexity. The selector is a small MLP with input dimension 12, hidden layer size 4, and 3 logits over $V_t \in \mathbb{R}^{HW \times D},$6. Training labels are derived offline by running the agent with each candidate structure and combining a judge reward and a memory utilization reward: $V_t \in \mathbb{R}^{HW \times D},$7 with $V_t \in \mathbb{R}^{HW \times D},$8 and $V_t \in \mathbb{R}^{HW \times D},$9.

A second distinctive component is the Beta Mixture Model (BMM)–based probabilistic gate for deciding whether new content should merge into an existing episodic session. Rather than thresholding raw similarity, the method normalizes scores to $\mathcal{M}^s$0, fits a two-component Beta mixture by EM, identifies the higher-mean component, and keeps candidates whose posterior gate value exceeds a threshold. In the main experiments, the posterior threshold is $\mathcal{M}^s$1 and the minimum keep is $\mathcal{M}^s$2. This replaces brittle similarity thresholds with a distribution-aware probabilistic criterion.

The reported experimental gains are benchmark-specific and substantial. On PERSONAMEM, FluxMem achieves 72.43% average accuracy, compared with a best baseline average of 63.25%, corresponding to an average improvement of +9.18% (Lu et al., 15 Feb 2026). On LoCoMo, it reports average F1 51.16, BLEU‑1 41.73, and ROUGE‑L 49.51, with improvements of +6.84, +2.10, and +9.49 over the best baseline. The backbone LLM is GPT‑4.1 (temperature 0), and retrieval combines all‑MiniLM‑L6‑v2, BM25, and reciprocal rank fusion.

The conceptual significance of this variant is that “FluxMem” no longer means token compression or buffer management. It means memory-structure selection as an explicit learned control variable. This sharply distinguishes it from the video variant, which is fully training-free.

4. Continuously evolving connectivity in agent memory

In "Rethinking Memory as Continuously Evolving Connectivity" (Fang et al., 27 May 2026), FluxMem is reformulated again, this time as a connectivity-evolving memory framework in which memory is a heterogeneous graph

$\mathcal{M}^s$3

rather than a static repository. The graph has three layers: semantic knowledge, episodic experiences, and procedural skills. Semantic nodes contain facts, documents, or API documentation; episodic nodes contain full trajectories

$\mathcal{M}^s$4

procedural nodes contain distilled reusable strategies. The principal cross-layer edge types are grounding edges from semantic to episodic memory and distillation edges from episodic to procedural memory.

The framework evolves through three stages. Stage I – Initial Connection Formation retrieves semantic nodes with a hybrid score combining cosine similarity, BM25, and LLM verification: $\mathcal{M}^s$5 retrieves episodic nodes by embedding similarity, and inherits procedural skills through distillation edges. Stage II – Feedback-Driven Connectivity Refinement edits the active subgraph online: it performs link expansion for under-connection, link pruning for over-connection, and unit-level refinement when a memory unit is too coarse or too fine for the current task. Stage III – Long-Term Connection Consolidation clusters episodic trajectories and induces procedural skills, then iteratively refines them with the Procedure Evolution Maturity Score (PEMS): $\mathcal{M}^s$6

This FluxMem is therefore not a selector over fixed structures, but a framework in which connectivity itself is the optimization target. The paper emphasizes that context at time step $\mathcal{M}^s$7 is an activated subgraph serialized into the prompt: $\mathcal{M}^s$8 The principal failure modes are likewise graph-theoretic: under-connection, over-connection, and abstraction misalignment of memory units.

The reported benchmark results are state-of-the-art across three distinct settings. On LoCoMo, with GPT‑4.1‑mini, FluxMem reaches 95.06 average LMJ, compared with 93.05 for EverMemOS and 81.23 for a Full Context baseline (Fang et al., 27 May 2026). With Qwen3‑30B‑A3B, it reaches 93.44. On Mind2Web, under the realistic setting, GPT‑4.1‑mini achieves Cross-Task SR 8.1 with FluxMem, compared with 3.6 for AWM and 2.8 for No Memory; Gemini‑2.5‑flash reaches 9.6. On GAIA, FluxMem reaches 64.85 average with Kimi K2, 76.36 with GPT‑5‑mini, and 70.30 with DeepSeek V3.2.

The ablations are important for interpretation. On LoCoMo, removing Stage II yields the largest drop, from 95.06 to 85.32 for GPT‑4.1‑mini. Increasing Stage II refinement depth from 0 to 5 improves LoCoMo performance from 85.32% to 95.06%. With Stage III active, PEMS increases from 0.072 to 0.158 in the first four rounds and essentially converges by round 5. This supports the paper’s claim that memory should be modeled as continuously evolving connectivity rather than static retrieval.

5. Flux-controlled memory in superconducting and bosonic systems

Outside AI memory architectures, the term appears in a looser but technically relevant sense for flux-controlled physical memory systems. In "Hysteretic superconducting heat-flux quantum modulator" (Guarcello et al., 2017), a thermally biased DC SQUID with finite loop inductance exhibits hysteresis in both thermal current and the temperature of one electrode. The abstract states that the proposed device can effectively find application as a temperature-based superconducting memory element, with temperature jumps up to, e.g., ~ 38mK for a realistic Al-based setup, working even at GHz frequencies by suitably choosing the superconductor. The detailed exposition identifies two thermal states associated with different fluxoid numbers $\mathcal{M}^s$9, with an explicit example of a hysteretic thermal jump $\mathcal{M}^m$0, interpreted as distinguishable “heat-bit 1” and “heat-bit 0” states.

A second hardware-related usage appears in "Flux-Activated Resonant Control of a Bosonic Quantum Memory" (Valadares et al., 20 Feb 2026). Here the central system is a long-lived bosonic memory in a 3D superconducting cavity, integrated with an on-chip flux-control architecture that dynamically accesses resonant Jaynes-Cummings (JC) interactions. The memory mode frequency is

$\mathcal{M}^m$1

with measured cavity lifetime

$\mathcal{M}^m$2

and

$\mathcal{M}^m$3

The transmon ancilla is flux-tunable over

$\mathcal{M}^m$4

with coupling

$\mathcal{M}^m$5

By bringing the transmon into resonance, the experiment realizes efficient arbitrary rotations between any pair of Fock levels. In the resonant regime,

$\mathcal{M}^m$6

with JC eigenstates $\mathcal{M}^m$7 and eigenfrequencies $\mathcal{M}^m$8. The work demonstrates Fock-state preparation, multitone control, and Givens rotations between arbitrary Fock levels, including a rotation between $\mathcal{M}^m$9 and $\mathcal{M}^l$0.

These hardware papers do not define a common software framework called FluxMem. Rather, they show that in superconducting-memory discourse the label can denote, or be used to motivate, flux-controlled memory elements whose state is encoded thermally or bosonically. This suggests a domain-specific extension of the term from adaptive symbolic or neural memory to controllable physical memory substrates.

6. Comparative interpretation

Across these usages, FluxMem consistently denotes a mechanism for selective retention under constrained context, but the operative object differs sharply by field. In streaming video understanding, the object is the visual token stream, and the governing operations are temporal pruning and spatial consolidation (Xie et al., 2 Mar 2026). In adaptive LLM-agent memory, the object is the episodic store, and the governing operation is learned structure selection among linear, graph, and hierarchical forms (Lu et al., 15 Feb 2026). In connectivity-evolving agent memory, the object is the heterogeneous memory graph, and the governing operations are rewiring, pruning, and procedural consolidation under feedback (Fang et al., 27 May 2026). In superconducting systems, the object is a physical state variable—temperature, fluxoid number, or bosonic Fock excitation—controlled by magnetic flux or flux-activated coupling (Guarcello et al., 2017, Valadares et al., 20 Feb 2026).

Another common misconception is that all FluxMem systems are training-free. This is false in the literature. The video system explicitly does not introduce new learnable parameters and does not require fine-tuning (Xie et al., 2 Mar 2026). By contrast, the LLM-agent structure-selection system explicitly trains a selector with cross-entropy over offline labels derived from downstream rewards (Lu et al., 15 Feb 2026). The connectivity-evolving framework does not present itself as a training-free token manager; it relies on iterative LLM-mediated refinement and offline consolidation (Fang et al., 27 May 2026).

A further distinction concerns the meaning of “hierarchy.” In the video system, hierarchy is a memory-age hierarchy—short-term, mid-term, and long-term—coupled to token compression strength (Xie et al., 2 Mar 2026). In the adaptive agent-memory system, hierarchy is both memory-level separation and one candidate internal memory structure within MTEM (Lu et al., 15 Feb 2026). In the connectivity-evolving framework, the hierarchy is a three-layer heterogeneous graph over semantic, episodic, and procedural memories (Fang et al., 27 May 2026). The shared term therefore masks materially different formal objects.

Taken together, the literature indicates that FluxMem has become a broad label for memory systems that reject static, one-pass storage in favor of adaptive organization, progressive consolidation, or controllable state evolution. What unifies the name is not a single algorithmic core, but a research orientation: memory is treated as an active substrate whose form, granularity, or dynamics must be engineered in response to task structure, long-horizon reasoning demands, or physical control constraints.