Progressive Memory Management

• Progressive Memory Management is a design pattern that treats memory as an evolving resource, dynamically updating, compressing, and reorganizing content based on utility and constraints.
• It applies techniques such as active curation, staged freezing, and adaptive migration across LLM agents, federated training, and system memory hierarchies to optimize performance.
• By balancing retention with computational overhead, this approach improves efficiency, reduces latency, and supports scalable, long-running applications in diverse domains.

Progressive memory management denotes a family of techniques that treat memory as an evolving resource rather than a fixed buffer. Across recent work, this evolution takes several forms: active curation of external memory for LLMs, progressive compression of dialogue and affective histories, value-driven promotion and eviction in long-running agent systems, staged freezing of model blocks during federated training, and adaptive migration of pages across heterogeneous hardware tiers (An, 8 Aug 2025, Bamidele, 22 Feb 2026, Wu et al., 2024, Liu et al., 2017). This suggests that the term is best understood as a cross-domain design pattern: memory is updated, compressed, reorganized, or restricted over time so that the effective working set remains aligned with utility, task state, and system constraints.

1. Scope, terminology, and recurring abstractions

The phrase has distinct meanings in different literatures. In LLM and agent systems, “progressive” typically refers to staged retention, summarization, retrieval, forgetting, or compression over an interaction history or task trajectory (An, 8 Aug 2025, Yu et al., 5 Jan 2026, Gao et al., 26 Jun 2026). In federated learning, it refers to training the model “in a progressive manner” by dividing it into blocks and freezing converged blocks so that only a subset of parameters requires backward computation at each stage (Wu et al., 2024, Yebo et al., 2024, Wu et al., 2024). In operating systems and memory hierarchies, it refers to dynamic placement and migration across tiers such as DRAM, NVM, CXL-attached memory, cache, channels, and banks (Liu et al., 2017, Oren, 2017, Kadekodi et al., 26 Oct 2025). By contrast, in transactional memory, “progressive” is a liveness guarantee: a transaction is only allowed to abort if it actually conflicts with a concurrent transaction (Kuznetsov et al., 2015).

A compact way to organize the field is by the object being managed and the timescale of adaptation.

Scope	Characteristic mechanism	Representative papers
LLM agents and dialogue	Active curation, summarize/discard, retrieval, memory tools	(An, 8 Aug 2025, Yu et al., 5 Jan 2026, Ye et al., 11 Feb 2026, Gao et al., 26 Jun 2026, Huang et al., 2023)
Federated training	Block partitioning, freezing, output modules	(Wu et al., 2024, Yebo et al., 2024, Wu et al., 2024)
Systems and memory hierarchies	Promotion/demotion, page migration, hotness tracking, lifecycle tiers	(Liu et al., 2017, Oren, 2017, Kadekodi et al., 26 Oct 2025, Zheng et al., 13 Nov 2025, Bamidele, 22 Feb 2026)
Multimodal temporal reasoning	Progressive attention, adaptive memory buffers	(Kim et al., 2019, Jin et al., 4 Dec 2025)

Despite the heterogeneity of these settings, the recurring abstractions are similar: a bounded working set, a policy for retaining or discarding state, and a mechanism for estimating future usefulness. This suggests that progressive memory management is less a single algorithm than a systems principle.

2. Active and unified memory in LLM agents

A prominent line of work argues that memory management for LLMs should be task-driven and agentic rather than passive. Cognitive Workspace explicitly opposes the assumption that simply extending context windows solves the memory problem. It introduces active memory management with deliberate information curation, hierarchical cognitive buffers, and task-driven context optimization; its metacognitive controllers “continuously assess information relevance, predict future needs, and reorganize the workspace,” maintain priority hierarchies, introduce forgetting curves for stale information, and consolidate frequently used content into compressed forms (An, 8 Aug 2025). Its buffer hierarchy is specified as an Immediate Scratchpad (8K tokens), Task Buffer (64K), Episodic Cache (256K), and Semantic Bridge (1M+), with distinct retention policies, update strategies, and roles (An, 8 Aug 2025).

The empirical claims in that work are unusually explicit. Across multi-turn dialogues, extended sessions, multi-hop reasoning, and conflict resolution, Cognitive Workspace reports a memory reuse rate of 54–60% across tasks, averaging 58.6%, whereas classic RAG is reported at 0% (An, 8 Aug 2025). Despite a 3.3x increase in memory operations, net efficiency improved by 17–18%, with the paper defining

$$\eta = \text{reuse\_rate} / (1 + \text{extra\_operation\_ratio})$$

and reporting $p < 0.001$ together with Cohen’s $d$ from 23 up to 196 (An, 8 Aug 2025). It further characterizes operation growth as sublinear, $O(\log n)$, versus linear $O(n)$ for RAG (An, 8 Aug 2025).

Other agent frameworks push the same theme toward end-to-end policy learning. Agentic Memory unifies long-term memory and short-term memory by exposing Add, Update, Delete, Retrieve, Summary, and Filter as tool-based actions within the agent’s policy, and trains the policy with a three-stage progressive reinforcement learning strategy plus step-wise GRPO (Yu et al., 5 Jan 2026). UMEM argues that memory extraction and memory management must be jointly optimized; it introduces Semantic Neighborhood Modeling and a neighborhood-level marginal utility reward via GRPO, and reports up to a 10.67% improvement in multi-turn interactive tasks together with a monotonic growth curve during continuous evolution (Ye et al., 11 Feb 2026). SWE-MeM, targeted at long-horizon coding agents, gives the agent a flexible memory tool that decides when, what, and how to compress, and reports 43.4% and 60.2% resolve rate with 4B and 30B models, respectively, on SWE-Bench Verified (Gao et al., 26 Jun 2026).

Taken together, these papers reject a purely reactive retrieval model. Their shared claim is that effective long-horizon reasoning depends on explicit policies for anticipation, summarization, deletion, and reuse, not only on larger raw context windows.

3. Transparent curation and progressive compression in conversational and companion systems

A separate but related tradition emphasizes visibility and user control. Memory Sandbox turns conversational history into explicit “memory objects” that users can view, manipulate, record, summarize, and share across conversations (Huang et al., 2023). Each memory object can be edited, added, deleted, rearranged, and toggled visible or hidden; the set of visible memory objects, as ordered and toggled by the user, defines the actual input context sent to the LLM at each prompt (Huang et al., 2023). The paper formalizes the active context as a dynamic subset $M \subseteq \{m_1, m_2, ..., m_n\}$, where each memory object $m_i$ has a visibility bit $v_i$, a possible position index $p_i$, and content $c_i$ (Huang et al., 2023). Summarization creates a summary memory object from selected contents, and sharing copies memory objects by reference across conversations, preserving provenance (Huang et al., 2023).

This work reframes progressive memory management as staged human-guided curation rather than autonomous optimization. The central claim is not that the system chooses the optimal memory state, but that transparency repairs conversational breakdowns by aligning the model’s effective context with the user’s mental model (Huang et al., 2023). A common misconception in LLM interface design is that memory should remain hidden and automatic; Memory Sandbox is an explicit counterexample.

Livia applies progressive memory compression to an emotion-aware AR companion. Its memory system is managed by a Memory Compression Agent and combines Temporal Binary Compression (TBC) with a Dynamic Importance Memory Filter (DIMF) (Xi et al., 12 Aug 2025). TBC organizes memory into exponentially increasing time intervals in which recent interactions are kept in full detail, older interactions are summarized, and distant past is reduced to high-level summaries; pairwise summarization and hierarchical merging progressively reduce granularity with age (Xi et al., 12 Aug 2025). DIMF assigns importance scores using emotional intensity, contextual uniqueness, and user feedback, and periodically prunes or further compresses low-importance entries while preserving meta-entries or higher-level summaries for coherence (Xi et al., 12 Aug 2025).

The storage and recall numbers are concrete: the study reports reduction from 50KB to 15KB per user, up to 70% storage reduction, and critical event recall at 92% (Xi et al., 12 Aug 2025). Here, progressive memory management is explicitly modeled after human memory decay: detail decreases with age, but salient events are retained.

4. Lifecycle, tiers, and memory as a systems resource

In long-running systems, the main problem is often not retention itself but the computational footprint of retained memory. AMV-L states this directly: age-based retention such as TTL bounds item lifetime but does not bound the request-path computational footprint, so retrieval candidate sets and vector similarity scans can grow unpredictably (Bamidele, 22 Feb 2026). Its answer is Adaptive Memory Value Lifecycle, which gives each memory item a continuously updated utility score,

$p < 0.001$0

and uses value-driven promotion, demotion, and eviction across hot, warm, and cold tiers (Bamidele, 22 Feb 2026). Retrieval is restricted to a bounded, tier-aware candidate set, specifically hot tier plus a bounded sample from warm tier (Bamidele, 22 Feb 2026).

The serving-system results are strong and narrowly focused on latency tails. Relative to TTL, AMV-L improves throughput by 3.1x and reduces latency by 4.2x at median, 4.7x at p95, and 4.4x at p99, while reducing requests exceeding 2s from 13.8% to 0.007% (Bamidele, 22 Feb 2026). Relative to LRU, it accepts a small regression in median and p95 latency, reported as +26% and +3%, in exchange for improved p99, reported as -15%, a -98% reduction in requests exceeding 2s, approximately 6% fewer tokens per request, and retrieval value means within approximately 0–2% (Bamidele, 22 Feb 2026). The paper’s broader argument is that predictable performance requires explicit control of working-set size rather than retention time alone.

Operating-system and architecture papers instantiate the same idea in physical memory hierarchies. memos schedules resources across the entire memory hierarchy, including cache, channels, and DRAM/NVM main memory, using a kernel-level monitoring module, page migration engine, page-coloring, and predictive placement policies (Liu et al., 2017). Its reported averages are 19.1% system throughput improvement, 23.6% QoS improvement, 3–83.3% lower NVM-side latency, 25.1–99% lower energy consumption, and 40X average NVM lifetime improvement (Liu et al., 2017). The multi-level Aging work likewise argues that manual Memory Allocation Manager control is optimal in theory but unrealistic in HPC practice, and that adapting paging algorithms from two-level to $p < 0.001$1-level hierarchies is the practical route; it reports that the Aging adaptation gives the best performances in terms of Hit/Miss ratio and that, with 3 memory levels, the 3-level Aging algorithm gave up to 3× the hit ratio than a one-level approach (Oren, 2017).

Jenga targets DRAM plus capacity-tier memory and addresses two specific failure modes: hot and cold objects placed in the same page, and abrupt hotness measurements that lead to thrashing (Kadekodi et al., 26 Oct 2025). It groups heap objects by allocation context, using return addresses of up to 10 stack frames to assign objects to one of 32 allocation regions, and smooths page hotness decay so that the system reacts in a timely manner while avoiding thrashing (Kadekodi et al., 26 Oct 2025). Compared to the best previous tiered memory system, it runs memory-intensive applications 28% faster across 10 applications, with CPU overhead of less than 3% of a single core and memory overhead of less than 0.3% (Kadekodi et al., 26 Oct 2025).

Vmem addresses a different systems requirement: hot upgradability in production cloud environments. Its architecture separates a stable interface module, vmem.ko, from an online-upgradable logic module, vmem_mm.ko, and experiments report about 2% higher sellable memory rate, over 3x faster boot time for VFIO-based VMs, and about 10% better network performance for DPU-accelerated VMs; the system is reported as deployed for seven years on over 300,000 cloud servers supporting hundreds of millions of VMs (Zheng et al., 13 Nov 2025). Across these systems papers, progressive memory management appears as lifecycle control over physical or logical working sets, with explicit attention to overhead, latency, and upgradeability.

5. Progressive training under memory constraints

A substantial body of work uses “progressive” to describe memory-efficient training itself. ProFL partitions the model into sequential blocks, trains them one at a time, freezes converged blocks, and divides the process into model shrinking and model growing stages so that each block can learn suitable feature representations with corresponding output modules (Wu et al., 2024). It proposes a scalar-based effective movement metric to decide when a block is safe to freeze and reports up to 57.4% reduction in peak memory footprint together with up to 82.4% improvement in model accuracy (Wu et al., 2024).

SmartFreeze follows the same block-wise pattern but adds a pace controller and a participant selector. The pace controller monitors training progress of each block at runtime and safely freezes it after convergence, while the participant selector chooses devices by jointly considering memory capacity together with statistical and system heterogeneity (Yebo et al., 2024). The reported gains are up to 82% average memory reduction, up to 83.1% model-accuracy improvement, and up to 2.02X acceleration (Yebo et al., 2024).

NeuLite introduces elastic progressive training, a Curriculum Mentor, and a Training Harmonizer. Instead of updating the full model throughout training, it divides the model into blocks and trains them progressively; curriculum-aware losses are designed for each block, while parameter co-adaptation and output modules are used to reduce information isolation across blocks (Wu et al., 2024). It reports up to 50.4% reduction in peak memory usage, up to 84.2% improvement in model performance, and up to 1.9X faster training (Wu et al., 2024).

These papers define progressive memory management at training time rather than inference time. The underlying mechanism is simple but consequential: by freezing blocks, the system avoids backward computation and storage of activations and gradients for frozen parts of the network. A plausible implication is that “memory management” in this literature includes not only the retention of semantic content, but also the scheduling of gradient-bearing state.

6. Temporal pruning, multimodal memory, and design tensions

Progressive memory management is also a strategy for temporal localization in multimodal reasoning. The Progressive Attention Memory Network for movie story question answering uses dual memories for video and subtitle streams and applies progressive attention in multiple steps, first with question cues and then with answer cues, so that irrelevant temporal parts are progressively down-weighted (Kim et al., 2019). On MovieQA, it reports 42.53% test accuracy; the ablation “PAMN w/o PA” drops validation accuracy from 43.34% to 42.03%, removing multiple hops drops to 42.67%, and per-question-type analysis reports gains of 13–20% on “where” and “when” questions (Kim et al., 2019). Here, progressive memory management is essentially progressive pruning of a temporal evidence store.

VideoMem generalizes this principle to ultra-long video understanding. It models the problem as sequential generation with an adaptive global memory buffer that retains critical information and discards redundant content across the video timeline, and trains with Progressive Grouped Relative Policy Optimization, combining Progressive State Propagation and Temporal Cascading Reward (Jin et al., 4 Dec 2025). The reported benchmark improvements over the Qwen3-VL-8B base are +5.7% on Long-duration VideoMME, +7.9% on LVBench, and +8.9% on LongTimeScope; removing TCR lowers performance by 4.2 to 4.9%, removing PSP lowers it by 2.1%, and PRPO is reported as 3.1× faster than vanilla GRPO with convergence in 70% of steps (Jin et al., 4 Dec 2025).

Several design tensions recur across the literature. One is the difference between passive accumulation and active curation: Cognitive Workspace argues that Infini-attention and StreamingLLM can handle longer contexts but cannot decide what is relevant or strategically forget (An, 8 Aug 2025). A second is the difference between retention and bounded compute: AMV-L shows that TTL can preserve memory while still producing heavy-tailed latency because eligibility for retrieval remains unbounded (Bamidele, 22 Feb 2026). A third is the trade-off between manual and automated control: multi-level hierarchy work treats manual allocation as theoretically optimal but practically unrealistic, whereas Memory Sandbox deliberately reintroduces user control at the interface level, not at the OS page level (Oren, 2017, Huang et al., 2023). Finally, the term “progressive” itself is polysemous: in transactional memory it names a conflict-based commit guarantee, whereas in contemporary LLM, FL, and systems papers it names staged or adaptive memory evolution (Kuznetsov et al., 2015).

The literature therefore converges less on a single canonical architecture than on a shared thesis: memory should be managed according to evolving value, not merely stored. Whether the unit is a prompt span, a memory object, a model block, a page, a VM slice, or a temporal segment, progressive memory management treats memory state as something to be continuously curated, rather than passively accumulated.