NEMORI: Replacing Designer Intuition About What to Remember
with the System's Own Prediction Error

0 · Search Disclosure

• Search scope: ACL Anthology 2026 (acl-long / findings-acl / eacl-long), arXiv cs.CL / cs.AI, OpenReview (ICLR 2026 MemAgents workshop), Agent-Memory-Paper-List.
• Candidate pool: Scanned 26 papers covering eight subfields: memory architecture, distillation, compression, evolution, RL-trained memory policies, evaluation benchmarks, security and governance, and parametric memory.
• Exclusions: As agreed, MemGPT, A-MEM, Mem0, Memory-R1, MemAct, and AgeMem were excluded as core papers. arXiv:2607.08032 (rate–distortion memory compression) was also excluded because its formal framework had already appeared as an example in an earlier report.
• Transparency statement: tabzhangjx.github.io/year-archive/ returned a retrieval-permission error in this run; the historical topic list could not be checked programmatically, so deduplication relied only on the exclusions in the task prompt.
• Selection rationale: An ACL 2026 main-conference long paper in the previously unvisited memory-distillation/compression rotation, with complete, verifiable quantitative matrices: main results, costs, ablations, retrieval strategies, third-party integration, scalability, and two hyperparameter sensitivity analyses.

2.1 Abstract Summary and Positioning

NEMORI's problem fits in one sentence: How should a memory system decide which information is worth retaining? The authors diagnose existing answers as encoded designer intuition: importance scores in Generative Agents, emotion tags in emotional memory, fact templates in Mem0, and heat scores in MemoryOS. The paper describes these methods as “encoding designer intuition rather than learning from the data itself”, with two consequences: subjective bias causes irreversible information distortion, while systemic bloat causes excessive storage and amplifies retrieval noise.

NEMORI draws on predictive coding (Rao & Ballard 1999; Friston's 2010 free-energy principle): information predictable from existing knowledge is redundant; the unpredictable portion is the candidate for memory. The system first uses existing semantic knowledge to generate an anticipatory schema for an upcoming interaction segment, then distills the difference between actual content and expectation into semantic knowledge.

2.2 Motivation and Problem Evidence

The argument develops through three levels.

Level 1: Capacity Constraints Are Hard Limits. Stateless LLMs rely on growing interaction histories despite limited context windows and the Lost in the Middle phenomenon. This is field-wide consensus rather than NEMORI's distinctive claim.

Level 2: Heuristic Distillation Introduces Irreversible Loss. This is the paper's real target. Table 1 reorganizes memory systems around Distillation / Management / Retrieval three stages (see Table A), revealing a structural fact: only five of 13 baselines have a distillation entry, and all use manually imposed priors; another seven leave that column blank, storing raw interactions wholesale and shifting the burden to management and retrieval.

Level 3: The Problem Is Quantified.

• Ablation (Table 5): Replacing prediction-error distillation with direct distillation (Nemori-s) changes the average LoCoMo LLM-judge score from 73.0 → 51.7 (gpt-4o-mini, −29.2%, calculated in this report), 80.8 → 66.0 (gpt-4.1-mini, −18.3%). Changing only the distillation criterion within the same episodic pipeline creates a relative gap of nearly 30%.
• Construction cost (Table 3): NEMORI's total memory-construction token count is 322.9k, versus Mem0's 1693.4k, A-MEM 1149.4k, LangMem 1010.2k. NEMORI reduces cost by 80.9%(relative to Mem0, calculated here) while scoring higher (73.0 vs 61.3).

2.3 The Paper's Argument

The argument is not simply a list of modules. It is a deductive chain linked by three priors that form the paper's backbone and distinguish it from another generic memory framework.

2.4 Methodology and Mechanisms

Table A · Paper Table 1: Memory Systems Reorganized by Distillation / Management / Retrieval (Transcribed)

Formal Definition · Module 1: Episodic Memory Integration

Formal Definition · Module 2: Semantic Knowledge Distillation (Core Contribution)

Table B · Hyperparameters and Sensitivity Coverage

Pipeline and Cost Breakdown

2.5 Experimental Design and Quantitative Results

Evaluation Setup

Table C · Main LoCoMo Results (Paper Table 2; LLM-Judge / F1 / BLEU-1)

gpt-4.1-mini

gpt-4o-mini

Table D · Independently calculated: NEMORI vs Full Context category-level results (LLM-judge)

Table E · Memory-Construction Cost (Paper Table 3; LoCoMo / gpt-4o-mini)

* Cost per point is calculated here as total tokens divided by the LLM-judge score, comparing efficiency after accounting for performance differences. NEMORI's cost per score point is 1/6.3 of Mem0's and 1/2.2 of MemoryOS's. The paper's ↓38.7% compares against the cheapest baseline, MemoryOS; relative to Mem0, the reduction is ↓80.9%.

Table F · Ablations (Paper Table 5) + Relative Declines Calculated Here

Table G · Construction-Cost Breakdown (Paper Table 9, gpt-4o-mini)

Table H · Retrieval-Strategy Ablation (Paper Table 6, LoCoMo / gpt-4.1-mini)

Across four combinations, the LLM-judge range is only 1.7 pp(75.3–77.0), and the F1 range is 0.7 pp. This ablation actually suggests that indexing/retrieval representation barely matters, in tension with the strong Representation Prior in §2.3. N→N is the default, yet N→P achieves a slightly higher LLM-judge score of 77.0.

Table I · Third-Party Integration (Paper Table 7): NEMORI as a Front-End Distiller

Table J · Scalability (Paper Table 8, LongMemEval-S, 105K Average Tokens)

Full Context uses 101K tokens; NEMORI uses 3.7–4.8K (↓95–96%). Relative changes are calculated in this report.

Table K · Hyperparameter Sensitivity

Observation window w (Table 11, gpt-4.1-mini)

Retrieval count k (Figure 3 / Table 12)

* Extracted Figure 3 and Table 12 values differ slightly at k=2 and k=30 (65.0 vs 66.6; 73.0 vs 74.4). This report could not determine which values reflect the main-text convention and presents both transparently.

2.6 A Phase Transition: The Most Convincing Qualitative Evidence

This goes beyond a marginal compression gain; it is a sign reversal: Reducing context to 1/25 improves performance by 5–7×. For preference questions, raw context is actively harmful rather than merely insufficient—preferences are scattered across long conversations and buried in irrelevant content, preventing localization within 101K tokens. Prediction-error distillation isolates new preferences inconsistent with the existing profile as semantic knowledge.

This is the paper's only evidence of a change in kind rather than degree : Elsewhere, differences from Full Context remain within ±20% (Table J); Preference alone approaches an order-of-magnitude change. If the paper were to center one claim, it should be that memory combats context dilution rather than context length, but it does not develop this transition as a standalone core argument.

The Counterevidence Also Deserves Attention: Knowledge Update sees gpt-4o-mini fall from 78.2 to 61.5 (−21.4%). Yet conflict is one of NEMORI's three semantic-consolidation states, specifically designed for knowledge updates. The dedicated mechanism fails most severely on its intended question type—an internal contradiction the paper does not discuss.

2.7 Assessment of Value

The substantive contribution consists of three things beyond better performance:

1. Redefining what to store as a measurable quantity. Before NEMORI, importance, heat, emotion were all exogenous, requiring a prompt, threshold, or decay coefficient. Prediction error is endogenous, jointly determined by current knowledge and input, without an additional configuration interface. The shift resembles moving from handcrafted features to a self-supervised objective.
2. An existence proof of 80.9% lower cost and 19.1% better performance(Table E, relative to Mem0). It challenges the assumption that memory-system cost is unavoidable: much of it comes from storing unnecessary material.
3. A Detachable, Modular Contribution. Table I shows prediction-error distillation can serve as a front-end filter for other systems: A-MEM's storage falls 51–64% while Core Score rises 6.1%. The mechanism therefore does not depend on the rest of NEMORI and can be adopted directly by A-MEM, MemoryOS, or Mem0.

Independently Reusable Design Principles

2.8 Limitations

(A) Author-Acknowledged Limitations — Quoted from the Formal ACL Version

1. Distillation-centered scope: “NEMORI focuses on distillation and adopts naive strategies for management and retrieval. This simplicity suffices for the benchmarks studied here, but may become a bottleneck for tasks that demand more sophisticated reasoning over memory. The performance we report should be understood as reflecting this design scope rather than the ceiling of what a complete memory system could achieve.”
2. Interfaces are not yet standardized: “The interfaces we define are currently conceptual, and due to the lack of standardized protocols in the area, concrete integration still requires case-by-case implementation.”

(B) Nine Additional Criticisms, Each with Evidence

① The Near-Zero Advantage Over Full Context Is Obscured by the Reporting Baseline(Argument gap). All relative gains use the strongest memory-system baseline as denominator. Against Full Context (Table D), gpt-4.1-mini gains only 0.2pp in Average (+0.25%), while F1 is 1.2pp lower (52.1 vs 53.3, −2.25%) and Multi-Hop is 2.4pp lower. This differs from the abstract's impression of improved performance alongside reduced storage: when the full context fits, NEMORI's performance gains are within the noise scale, leaving cost as the real selling point. A narrower claim of equal performance at one-fifth the cost would be stronger.

② The Core Mechanism's Contribution Declines Sharply with Model Capability(Metric validity + extrapolation risk). In Table F, w/o s(removing the semantic layer distilled from prediction errors) reduces gpt-4o-mini by −25.1% but gpt-4.1-mini by only −4.8%. Across just one model generation, the marginal contribution shrinks 5.2×. Linear extrapolation to frontier models could imply a near-zero semantic-layer contribution: prediction-error distillation may compensate for weaker models' retrieval and integration abilities rather than supply irreplaceable information. No stronger model tests whether this trend levels off; this is the most needed experiment.

③ The Knowledge-Update conflict Mechanism Fails Most on Knowledge-Update Questions(Internal contradiction). Table J: Knowledge Update gpt-4o-mini 78.2 → 61.5 (−21.4%), the only double-digit negative change among six categories, while Preference improves 597% on the same model. If conflict really removes outdated entries, this should be its strongest category. A plausible hypothesis is that conflict depends on Search(D_s, u_q, K_m=5) retrieving the old entry. If old and new expressions fall below the similarity threshold, the conflict is never detected—pointing directly to the untested sensitivity of τ = 0.70. The paper offers no discussion.

④ The Critical Hyperparameter τ = 0.70 Has No Sensitivity Analysis(Experimental gap). Table B shows sweeps of w(5→40) and k(2→30), but none for τ. Yet τ Yes is the context-evocation threshold for anticipatory schemas, directly controlling how much the system believes it knows and thus prediction-error magnitude—the central knob of the Distillation Prior. Skipping τ while sweeping w examines secondary parameters while leaving the central mechanism unchecked. Additionally, K_e = K_m = 5, K_s = 10, r = 2 also lack ablations.

⑤ Native Management Has a Negative Contribution in Both Comparisons(Argument gap). Table F: Nemori-s 51.7 with management versus 52.0 without (−0.3); w/o e 64.6 with management versus 65.0 with naive RAG (−0.4). Management slightly hurts performance in both comparisons. The paper calls this negligible management overhead. More rigorously, single-run differences of ±0.3–0.4pp without variance cannot establish either harmlessness or harm, but the two differences have the same sign. Thus “Agnostic” in Table 1's Management column is more than a scope restriction: it reflects an unvalidated default.

⑥ No Head-to-Head Comparison with Learned Memory Policies(Missing key baselines). Another major 2025–2026 line delegates memory operations to RL-trained policies—Memory-R1, MemAct, and AgeMem—using exogenous rewards, whereas NEMORI uses endogenous prediction error. These directly compete on what should drive memory writes, yet among seven baselines, none comes from the RL branch. The implicit claim that endogenous signals outperform external rewards remains entirely untested. This is the most valuable next comparison.

⑦ Table I's Metrics Move in Opposite Directions, Raising Validity Questions(Metric issue). When NEMORI precedes A-MEM, LLM-judge falls 3.0%/3.9%, while Core Score rises 6.1%/6.1%. The paper uses the latter to claim maintained or improved performance. When two metrics disagree within one experiment, favoring the positive one requires justification. The paper does not explain Core Score's definition, its divergence from LLM-judge, or why the main table does not use it.

⑧ Narrow Evaluation Overpromises “LLM Agents” in the Title(Experimental design). LoCoMo and LongMemEval-S are both conversational question-answering benchmarks, with no tool use, web or embodied environments, or planning tasks such as ALFWorld, WebArena, or PDDL. Agent memory must retain behavioral experience—failed tool calls or unworkable paths—whose redundancy differs from conversational facts. LoCoMo also has only 10 conversations, 1,540 questions, one run, and no confidence intervals, while several key Table D differences are only 0.1–1.2pp, indistinguishable at this sample size.

⑨ Limitations of This Report (Disclosure). This analysis uses extracted text from the official ACL PDF and arXiv v4 HTML; the code was not run; github.com/nemori-ai/nemori was located through search only, without verifying whether it fully implements the paper. Its public description calls it “a minimalist MVP,” potentially differing from the paper's “a production-grade implementation”; reproduction efforts should note this. The discrepancy between Figure 3 and Table 12 also remains unresolved.

Selection criterion: All five works address the same question—what to remember or forget—and are explicitly classified in NEMORI's Table 1. They represent five ways to turn the same knob; NEMORI attempts to remove that knob.

3.1 Generative Agents — Park et al., UIST 2023

Problem addressed. The first operational long-term agent-memory loop: a memory stream records every observation, retrieval weights importance × recency × relevance three factors, and reflection synthesizes low-level observations into high-level insights. It is the first entry in NEMORI Table 1 with a populated Distillation column.

Unresolved problem.importance is rated 1–10 by a prompted LLM, without supervision linking that score to downstream utility. It proxies the designer's intuition about importance. Overrating creates systemic bloat; underrating causes irreversible loss. More fundamentally, importance is an absolute quantity that does not vary with existing knowledge: a sentence scores the same on its first and tenth occurrence.

How the core paper responds. NEMORI replaces an absolute quantity with a relative quantity: information value equals surprise relative to current knowledge. The tenth repetition is matched by Evoke , covered by the anticipatory schema, and yields no distilled output. Table F's Nemori-s drops 18.3–29.2%, quantifying the cost of replacing this criterion.

Transfer of key designs. Generative Agents' reflection tree evolves in NEMORI into the merge consolidation state, triggered only when semantically adjacent entries are retrieved rather than when accumulated importance crosses a fixed threshold. A counter becomes an associative match.

3.2 MemoryBank — Zhong et al., AAAI 2024

Problem addressed. Forgetting becomes a first-class operation: the Ebbinghaus forgetting curve makes memory strength decay exponentially with time and increase with repeated recall, providing an early computable answer to what to forget.

Unresolved problem. The forgetting curve is a prior about time rather than a content criterion. An unvisited week-old fact about a drug allergy decays at the same rate as unvisited week-old small talk. Importance is outsourced to access frequency, which depends on what questions are asked and provides little information during cold starts or infrequent use.

How the core paper responds. NEMORI avoids temporal decay. Forgetting happens before writing: predictable information never enters semantic storage, and stored entries are removed only during conflict when new information contradicts them. Time decay becomes predictability filtering plus contradiction-based removal—forgetting shifts from postprocessing to write gating.

3.3 A-MEM — Xu et al., 2025 (Brief by agreement; discussed in detail in earlier reports)

Problem addressed. Inspired by Zettelkasten, memory entries generate keywords and tags, link to existing entries, and trigger neighboring entries'memory evolution when new memories arrive. Without a predefined schema, it exemplifies self-organizing structure.

Unresolved problem. A-MEM optimizes organization among stored items but is silent on whether to store them: its Distillation column is empty. Links therefore form over unfiltered material; noisy entries also generate tags and links, propagating redundancy from storage into graph structure.

How the core paper responds. Table I answers directly: feeding NEMORI's distilled K rather than raw episodes P to A-MEM reduces MemTokens from 397K to 142K (↓64.3%), while Core Score rises from 52.6 to 55.8 (↑6.1%). A-MEM's organizational ability is unchanged; the material it organizes is cleaner.

Transfer of key designs. A-MEM's new-entry-driven updates become NEMORI's Supersede: M_ν direct replacement U_k, with P_k ‖ P_j preserving raw provenance. A-MEM uses append-style evolution(old entries are rewritten but remain), whereas NEMORI uses replacement-style evolution(old entries are displaced)—cheaper, but more irreversible.

3.4 MemoryOS — Kang et al., EMNLP 2025 Main (Oral)

Problem addressed. Operating-system paging becomes short-/mid-/long-term memory storage, using heat score to govern movement between tiers alongside two-stage retrieval. This is highly engineered heuristics; the paper reports a LoCoMo gain over its baseline of F1 +48.36%, BLEU-1 +46.18% (gpt-4o-mini).

Unresolved problem. Heat scores remain exogenous: migration thresholds, decay coefficients, and capacities are manually configured. More importantly, the system decides which tier to use, not whether to store an item—everything enters short-term storage. Hierarchy shifts the cost of bloat from retrieval latency to tier management.

Transfer of key designs. MemoryOS's hierarchy becomes two orthogonal stores in NEMORI (D_e episodic / D_s semantic), and migration shifts from heat thresholds to nonempty prediction errors. The hierarchy changes from time-based tiers to division by representation type, aligning with McClelland et al.'s Complementary Learning Systems theory, which the paper explicitly cites.

3.5 Zep / Graphiti — Rasmussen et al., 2025

Problem addressed. Use a temporal knowledge graph handles expiring memory: edges carry validity intervals, new facts invalidate rather than delete old edges, enabling point-in-time queries and conflict resolution.

Unresolved problem. Validity management first requires information to be structured as triples, itself a lossy distillation guided by manually chosen entity/relation schemas. Graph extraction is also expensive, while Zep averages only 58.5 / 61.6 on LoCoMo in the common evaluation; its Temporal Reasoning score of 60.2 (gpt-4.1-mini) trails NEMORI's 77.3—a system explicitly designed for temporal semantics loses by 17.1pp on temporal questions.

How the core paper responds. NEMORI uses conflict for the same problem at much lower cost: no graph maintenance, only deciding whether new evidence contradicts an old semantic entry. The price is losing point-in-time queries: NEMORI cannot answer what a user's preference was three months ago after purging that entry. Table J's Knowledge Update −21.4% may be the bill for this inexpensive conflict resolution.

Transfer of key designs. Zep's conservative invalidate-rather-than-delete strategy becomes purge-and-replace in NEMORI. This is the lineage's sole regression rather than advance: NEMORI is weaker on this dimension. Future work could restore temporal validity within the prediction-error framework (see §5 ⑤).

Underlying Logic: Criteria Become Increasingly Internal

Two Key Turning Points: IV→V delegates the criterion to optimization: if manual design fails, learn it from rewards, at the cost of task rewards and training. IV→VI delegates it to system state: let the system measure surprise, at the cost that criterion quality depends entirely on anticipatory-schema quality.

5.1 Theory

① Formalizing Cold-Start-to-Saturation Dynamics (High Value, Low Barrier). Prediction-error distillation output |K_in| is a function of existing knowledge |D_s| : |D_s| = 0 degenerates into full distillation, while |D_s| → ∞ approaches zero. The curve determines steady-state memory size, yet the paper never measures it. An actionable experiment: across LoCoMo's ten conversations, plot the joint trajectory of |K_in| with |D_s| by session index, testing whether saturation exists and depends on τ. This requires only logging the current pipeline, with no new training.

② τ and Prediction-Error Quality: Quantifying the Trade-Off.τ High ⇒ fewer evoked entries ⇒ weaker expectation ⇒ more false surprise (redundant storage); τ low ⇒ more evoked entries ⇒ stronger expectation ⇒ false negatives (missed new information). This is a standard precision–recall trade-off, but no empirical curve has been provided. Sweep τ ∈ [0.5, 0.9] and record (a) semantic-store size, (b) per-category LoCoMo scores, and (c) conflict trigger counts. The third directly tests the hypothesis in §2.8, criticism ③.

③ Comparing Attainable Limits of Endogenous Signals and External Rewards. Prediction error is an unsupervised criterion, whereas RL reward is a supervised criterion. In principle, the latter has no lower ceiling because it can learn any criterion, including prediction error. The question becomes: At what training budget does an RL memory policy surpass the training-free prediction-error criterion? This experiment could identify a crossover point and adjudicate the two branches in §4.

5.2 Engineering and Applications

④ Evaluate Agent Settings (Most Urgent). NEMORI's title says “for LLM Agents,” but both benchmarks are dialogue QA. On ALFWorld, WebArena, or τ-bench, define an episode as a tool-call sequence and prediction error as the difference between expected and actual tool outputs. The transfer is almost literal: Predictive coding naturally fits expected-versus-actual outcomes, precisely the structure of agent trajectories. Success here would be more persuasive than dialogue results.

⑤ Restore Zep's Temporal Validity Within Predictive Coding. NEMORI's conflict purges entries for efficiency, losing historical queries. Change conflict from purge-and-replace to marking invalid_at and appending, filtering by query time only during retrieval. Storage growth should be limited because old distilled entries are short, while temporal information is restored. Test directly on LongMemEval-S Knowledge Update and Temporal Reasoning to see whether this repairs the −21.4% regression.

⑥ Memory Stability in Online, Dynamic Settings. Current evaluation constructs memory offline and asks questions afterward; real agents read and write continuously. An unaddressed online risk is that early distillation errors contaminate anticipatory schemas, systematically distorting later prediction errors in a positive feedback loop. This closely resembles the How Memory Management Impacts LLM Agents(acl-long.27) finding of error propagation at ACL 2026—its use of future task outcomes as a memory-quality signal offers a ready-made way to clean NEMORI's semantic store retrospectively.

5.3 Emerging Directions

⑦ Reversible Distillation or Controllable Loss. NEMORI's distillation is irreversible: although M_j retains P_j the original text, the semantic layer K_in cannot be inverted. Could prediction-error distillation have an adjustable lossiness control, switching one memory system between expensive high fidelity and cheap low fidelity? This naturally connects to rate–distortion memory compression (arXiv:2607.08032): prediction error corresponds to residual coding in rate–distortion theory; NEMORI effectively implements a residual encoder without an explicit rate constraint. Making that constraint explicit is the shortest path to unifying the two lines.

⑧ Multiple Perspectives for Anticipatory Schemas. Currently, Evoke uses one similarity retrieval to generate one expectation. Generate K expectations from factual, preference, and causal perspectives; use the intersection of their errors as high-confidence new information and the union as a candidate pool. This may reduce the missed conflicts in §2.8, criticism ③. It parallels MAGMA's four orthogonal graphs (acl-long.1709), but at distillation rather than retrieval.

⑨ Prediction Error as a Memory-Poisoning Attack Surface. A concrete, as-yet-unproposed risk: an attacker who infers the semantic store could craft high-error but worthless inputs, forcing garbage into memory, because NEMORI's write gate depends entirely on manipulable surprise. This may be easier than attacking heuristics: the input need only look surprising, not important. Combined with arXiv:2607.06595's poisoning framework, this is an immediately testable attack.

Three Trends in This Table

1. Performance is saturating; cost is not. From Generative Agents to NEMORI, LoCoMo means rise to 73–81, but NEMORI exceeds Full Context by only 0.2–0.7pp (Table D). The next order-of-magnitude advance will not come from LoCoMo scores.
2. Self-reported gains and independent reproduction systematically diverge. MemoryOS (+48.4% → below Mem0 independently) and Zep (+18.5% → below Mem0 independently) are two separate examples.
3. Publication venues are moving upward. Many representative 2023–2025 works were arXiv papers or workshops (A-MEM still lacks a main-conference record; Zep lacks a publication record). By 2025–2026, the area regularly reaches EMNLP/ACL long papers and oral presentations, with tightening review standards.

How Memory Management Impacts LLM Agents: An Empirical Study of Experience-Following Behavior (2026-07, ACL 2026 Main Long)

• Systematically characterizes agents' experience-following behavior through error propagation with misaligned experience replay two failure modes, using future task outcomes as a quality signal for stored experience. Directly complementary to NEMORI: NEMORI gates writes; this work corrects quality afterward.
  ACL Anthology · PDF · Code: Not found

EverMemOS: A Self-Organizing Memory Operating System for Structured Long-Horizon Reasoning (2026-07, ACL 2026 Main Long)

• A neuroscience-inspired three-stage lifecycle—Episodic Trace Formation → Semantic Consolidation → Reconstructive Recollection—with MemCell/MemScene abstractions. Highly comparable to NEMORI's dual stores, but consolidation uses clustering rather than prediction error—an ideal head-to-head comparator.
  ACL Anthology · Code: Not found

MAGMA: A Multi-Graph based Agentic Memory Architecture for AI Agents (2026-07, ACL 2026 Main Long)

• Projects memories into semantic, temporal, causal, and entity graphs; retrieval becomes guided cross-graph traversal, addressing the entanglement of relations in a single similarity space. The multiperspective anticipatory schemas in §5.3 ⑧ could adopt this four-axis decomposition.
  ACL Anthology · Code: Not found

Dynamic Long Context Reasoning over Compressed Memory via End-to-End RL (LycheeMemory) (2026-07, ACL 2026 Main Long)

• Compressor / Gate / Reasoner plus end-to-end RL; context extrapolates from 7K to 1.75M tokens, with gains over MemAgent of 2× lower peak memory and 6× faster inference. This represents the latest engineering progress on §4's external-reward branch.
  ACL Anthology · Code: Not found

FOREVER: Forgetting Curve-Inspired Memory Replay for Language Model Continual Learning (2026-07, ACL 2026 Main Long)

• Aligns replay with model time, defined by optimizer-update magnitude, rather than training steps, across 0.6B–13B models. The parametric-memory counterpart to MemoryBank's forgetting-curve idea, allowing the two lines to be read together for the first time.
  ACL Anthology

Scaling Self-Evolving Agents via Parametric Memory (TMEM) (2026-06, arXiv)

• Uses online LoRA weight updates Δ_t to carry memory, allowing an agent to change behavior within a single episode. It shifts from looking up past experience to learning from it, outperforming summary- and retrieval-based baselines on LoCoMo, LongMemEval-S, and CL-Bench. A particularly important work on the parametric-versus-external-memory axis.
  arXiv:2606.04536

When Agents Remember Too Much: Memory Poisoning Attacks on LLM Agents (2026-07, arXiv)

• A systematic memory-poisoning attack framework. In connection with §5.3 ⑨, surprise-gated systems such as NEMORI may be easier to poison than heuristic systems because appearing surprising is easier than appearing important.
  arXiv:2607.06595

From Recall to Forgetting: Benchmarking Long-Term Memory for Personalized Agents (2026-04, arXiv)

• Unusually, it treats forgetting rather than recall as a primary evaluation target. LoCoMo and LongMemEval test only recall, structurally contributing to the small NEMORI–Full Context gap in Table D: Current benchmarks cannot reward correct forgetting.
  arXiv:2604.20006

8 · One-Sentence Summary

NEMORI's key contribution is not its LoCoMo score (+0.25% over Full Context), but evidence that 80% of memory-system cost goes toward information the system already knows—alongside the LongMemEval Preference transition from 6.7 to 46.7 when 101K tokens are reduced to 4K: the enemy of long context is dilution, not length.