OCR-Memory: Rendering Agent Memory as Images

Search Notes for This Issue

This issue covered ACL Anthology 2026 main track / Findings, EACL 2026, OpenReview (ICLR 2026 + MemAgents Workshop), and arXiv cs.CL/cs.AI, screening 20+ candidates. ACL 2026 main-track long-paper candidates included 2026.acl-long.27(the Experience-Following study), 2026.acl-long.474 (OCR-Memory), 2026.acl-long.1600 (GAM), 2026.acl-long.1607 (Adaptive Memory Distillation), 2026.acl-long.1709(MAGMA); EACL 2026 included H-MEM and Amory; Findings included RecMem.

The final selection was OCR-Memory for three reasons: (1) ACL 2026 main-track long paper, posted to arXiv in 2026-04 and sufficiently recent; (2) it opens the modality dimension of memory representation, whereas MemGPT / A-MEM / Mem0 / Memory-R1 / MemAct / AgeMem covered initially are all text-only orthogonal; (3) it first fully transfers the context optical compression lineage emerging in 2025Q4 into agent memory, with a clear technical genealogy.

Core paper

Bibliography: Jinze Li, Yang Zhang, Xin Yang, Jiayi Qu, Jinfeng Xu, Shuo Yang, Junhua Ding, Edith Cheuk-Han Ngai. OCR-Memory: Optical Context Retrieval for Long-Horizon Agent Memory. ACL 2026 (Volume 1: Long Papers), San Diego, California, July 2026. Affiliations: University of Hong Kong / University of North Texas / University of Tsukuba / Yonsei University.

Related Work Traced

Evaluation Benchmarks / Datasets

1.1 Abstract and Positioning

The paper describes OCR-Memory as using the visual modality as a high-density representation of agent experience, enabling arbitrarily long histories with minimal prompt overhead. It renders interaction histories into images with unique visual identifiers and retrieves through a locate-and-transcribe paradigm: visual anchors select relevant regions, then return the corresponding verbatim text, bypassing errors from generative retrieval.

1.2 Motivation and Problem Argument

The paper frames the problem as a three-way dilemma: full long-horizon trajectories are valuable but do not fit, and all three existing escape routes fail differently.

Quantitative Argument: Per-step inference context on Mind2Web goes from 3,980 → 596 tokens (6.7×)(Table 7); on NIAH, across 4k–32k, a stable 10.2×–10.7× compression ratio accompanies accuracy of 98.5 → 94.1 (Table 4).

Novelty: All three previous approaches trade off within text space and therefore lie on the same rate–distortion curve. OCR-Memory argues that changing representation space to pixels changes the curve—a modality-level jump rather than movement along one curve.

Why it is worth solving: All methods achieve only 2.2%–4.8% Task SR on Mind2Web (Table 1), leaving end-to-end long-horizon web-agent success near the floor. Context budgets are a direct constraint; the paper fixes the context window at 4096 tokens.

1.3 The Paper's Narrative and Argument Structure

The argument is not simply “we propose X,” but a four-step progression:

Step 1: Reframe “what to store” as “which representation to store it in.” If compression versus fidelity in text space is zero-sum, is there a representation naturally denser than token sequences? DeepSeek-OCR provides empirical motivation: a document page encoded with 100 visual tokens can beat GOT-OCR2.0 using 256 tokens.

Step 2 (the Key Turn): Acknowledge the New Visual Failure Mode, Then Bypass It. This is the paper's smartest move. Asking a VLM to decode text from rendered images directly inserts OCR errors into the agent's reasoning chain. DeepSeek-OCR's own curve shows that at compression <10×, accuracy is 97%, but at 20× it falls to about 60%. The authors infer: if decoding is unreliable, do not decode. The VLM's output is therefore limited to paragraph indices(0/1 labels), with text retrieved verbatim from disk.

Step 3: Turn Indexing into a Trainable Visual Task. Borrow Set-of-Mark: draw red boxes and numbers on rendered images, so the VLM need only identify relevant numbers. This turns open-ended generation into bounded multilabel classification that can be fine-tuned with weighted BCE and calibrated with softmax over “1”/“0” logits.

Step 4: Add Time, Automatically Demoting Older Memories. Keep the latest 5 steps at high resolution and the rest at low resolution. Once a low-resolution memory is retrieved, active recall restores high fidelity. This turns a forgetting curve into a resolution curve.

1.4 Methodology and Mechanisms

1.4.1 Formal Definitions

Memory Structure (Eq. 5)—note that text is not discarded.

Two-Stage Retrieval (Eq. 6)

Set-of-Mark Binary Relevance (Eq. 7–8) and Probability Calibration (Eq. 11)

Inference Threshold + Top-K Fallback (Eq. 12)

Age-Aware Multiresolution (Eq. 13) and Active Recall (Eq. 14–15)

Training Objective (Eq. 20)—weighted BCE, deliberately favoring recall.

1.4.2 Key Mechanism Comparison

1.4.3 Hyperparameters and Engineering Details (All from Appendix A)

1.4.4 Three-Stage Workflow

1.5.1 Experimental Setup

1.5.2 Main Results Matrix (Table 1) + Relative Gains

Absolute values are from Table 1; “Δ vs. runner-up” and relative gains were calculated here from that table.

Italics = column runner-up; highlighted row = best method.

1.5.3 Ablations

(a) Set-of-Mark Mechanism (Table 2)

This directly answers the counterfactual of replacing it with end-to-end OCR decoding.

Interpretation: Removing SoM for generative retrieval reduces Ele Acc by 7.3 pp and increases latency 3.1×—accuracy and efficiency collapse together, experimentally validating Step 2 of §1.3. Boxes without numbers still lose 4.6 pp, showing that discrete numbering itself, not visual segmentation, accounts for most of the gain: the model needs an identifiable discrete symbolic anchor, not merely a spatial region.

(b) Multiresolution Strategy (Table 3)

Interpretation: Dynamic uses 32% of the tokens to obtain 99.1% of Step SR(46.1 / 46.5). However, Table 1's headline 46.1 uses Dynamic; the paper's own best configuration is Static High-Res at 46.5. Reporting the slightly lower but cheaper configuration is methodologically transparent. It also means active-recall upsampling does not fully recover downsampling losses (0.4 pp residual); see §1.7-④.

1.5.4 Scaling and Compression (Table 4)

Key Observation: Compression ratio almost constant(10.2–10.7×, variation <5%). This suggests compression is a constant property of the renderer rather than adaptive to context length, in tension with the abstract's “retention of arbitrarily long histories” (§1.7-②). Accuracy decays approximately log-linearly, about −1.5 pp per doubling.

1.5.5 Backbone Generalization (Table 5)

Gains across the two backbones are stable or slightly larger(Qwen3-32B Ele Acc gain +13.4 pp > GPT-4's +12.5 pp). A plausible explanation: retrieval uses the same frozen 3B VLM, so evidence quality has greater marginal value for a weaker reasoning backbone.

1.5.6 Retrieval-Level Evaluation (Table 6)

Relative Recall@1 gain +49.1% is the paper's largest individual improvement—but also the figure requiring the greatest discount; see §1.7-③ on unequal comparisons.

1.5.7 System Overheads (Table 7)

1.5.8 🔺 Transition Highlight: Divergence Under Tight Token Budgets

This is a characteristic transition-like divergence: the curves may converge under generous budgets, although no data above 8192 are provided, but change qualitatively under tight ones. Text-RAG's smallest usable unit is a complete text passage; OCR-Memory's is a thumbnail of 64 visual tokens. Below the former's granularity floor, text-RAG does not merely deteriorate—it fails.

Disclosure: Figure 2's exact values are not tabulated, so this report can only describe its qualitative trend, not cite exact numbers. This is a limitation of the analysis.

1.6 Value Analysis

The real contribution is not storing memory as images, but three transferable design principles:

① Separate Representation from Responsibility: Vision Localizes; Text Supplies Content.
Why it matters: it separates a modality's Advantages(high density and spatial addressability) from its weaknesses(lossy decoding), using only the strengths. This principle transfers to any dense but lossy representation—latent memory, KV-cache compression, or vector-quantized indices. A representation that cannot reliably reconstruct compressed content can instead serve as an index rather than storage.

② Reduce Open-Ended Generation to Bounded Classification: SoM as a Retrieval Interface.
Why it matters: generative retrieval is hard to verify or calibrate and has high latency (Table 2: 5.3s). SoM turns it into K binary logits, making it fine-tunable, calibratable (Eq. 11), thresholdable (Eq. 12), and class-weightable (Eq. 20)—four engineering controls at once.

③ Forgetting Means Resolution Degradation, Not Deletion.
Why it matters: MemoryBank-style forgetting irreversibly; OCR-Memory's forgetting is lossy but reversible: active recall restores fidelity after downsampling. Conceptually, this resembles the distinction between retrieval failure and storage failure in human memory. The framework lets us transfer the idea to other layers: tiered KV-cache precision, embedding dimensions, or summary granularity.

1.7 Limitations

(A) Author-Reported Limitations (Original Quotations)

1. Training overhead:“unlike training-free retrieval baselines, our framework requires fine-tuning a specialized optical retrieval model, which incurs additional training resource overhead.”
2. Rendering and storage overhead:“the process of rendering interaction logs into images is computationally more expensive than direct text storage, and storing visual histories inevitably consumes more disk space than raw text logs.”
3. Deployment memory footprint:“deploying the system imposes an extra memory footprint, as the parameters of the vision encoder must be maintained in memory alongside the primary language model.”

(B) Additional Critique (Eight Independent Observations)

① [Metric Validity] The +0.5 pp Mind2Web Task SR Gain Likely Falls Within Noise.
OCR-Memory 4.8 vs. AWM 4.3, an absolute gap of 0.5 pp. The Mind2Web Cross-Task test set contains hundreds of tasks, so 0.5 pp corresponds to a single-digit number of tasks. The paper reports no random seeds, error bars, or significance tests. By contrast, the +4.7 pp Ele Acc gain is more persuasive. Conclusion: the four Mind2Web metrics presented equally in the abstract are not equally persuasive; Task SR deserves less weight.

② [Argument Gap] “Arbitrarily Long Histories” Exceeds the Evidence's 32k Limit.
The abstract claims “enabling retention of arbitrarily long histories,” but Table 4's scaling experiment ends at 32k, with accuracy already falling monotonically from 98.5 to 94.1 (−4.4 pp). Extrapolating −1.5 pp per doubling gives about 91% at 128k and 87% at 1M. More importantly, compression remains ~10.5× rather than adapting to length: each 10× history expansion also increases visual tokens 10×, with no mechanism preventing eventual context overflow. “Arbitrarily long” lacks evidentiary support; the accurate claim is “within 32k, near-constant 10× compression at an accuracy cost of <5 pp.”

③ [Missing Key Competitor] Table 6's Retrieval Comparison Is Unequal.
OCR-Memory's retrieval module is fine-tuned for 3 epochs on HotpotQA, while Dense Text-RAG is described as off-the-shelf dense retrieval without fine-tuning. Under this setup, Recall@1 rising from 52.7 to 78.6 (+49.1%) cannot be attributed to modality: it is confounded at least by fine-tuned versus zero-shot retrieval. Missing key experiment: fine-tune a comparably sized text retriever—E5 / ColBERTv2 / 3B text reranker—on the same HotpotQA data. Without this row, the central claim that visual indices outperform text indices is untested.

④ [Experimental Design] Active Recall Does Not Recover Downsampling Losses and Lacks a Direct Ablation.
Table 3: Dynamic 46.1 vs. Static High-Res 46.5, a residual of −0.4pp; Task SR 4.8 vs. 4.9, −0.1 pp. If active recall fully worked, restoring high resolution on hits, Dynamic should match High-Res. The persistent 0.4 pp gap implies restoration is too late or incomplete: memories can be missed because they are already low-resolution and thus never get the chance to be restored—self-fulfilling forgetting. The paper does not ablate “Dynamic without active recall”, so its contribution is unknown: it might raise 39.7 to 46.1, or add only 1 pp. This is the paper's most important missing ablation.

⑤ [Argument Gap] No Upper-Bound Control Lets a Long-Context LLM Read the Full Text.
All experiments hard-cap context at 4096 tokens. This is a 2023-era budget, while GPT-4-family and Qwen3-32B models supported roughly 128k context at writing time. The missing comparison is: with a 32k budget, does text-RAG still lose to OCR-Memory at 4k? Table 4 shows OCR-Memory accuracy falling to 94.1 at 32k, while 32k text-RAG performance is unknown. Without this upper bound, limited context is a motivation manufactured by experimental settings rather than demonstrated empirically.

⑥ [Argument Gap] Localization Requiring Much Less Information Than Reconstruction Is Never Directly Measured.
Directly test this by measuring, on the same memories, (i) low-resolution retrieval Recall and (ii) low-resolution OCR reconstruction accuracy. The paper provides only downstream metrics in Table 3 (39.7 Step SR), an indirect signal diluted by agent reasoning. Indirect evidence supports the assumption: DeepSeek-OCR reports about 60% OCR accuracy at 20× compression, while OCR-Memory at 512² still achieves 39.7 Step SR versus 37.9 zero-shot. Localization indeed does not collapse alongside reconstruction. But this is this report's inference, not the paper's proof.

⑦ [Experimental Design] Backbone Choice and Scale Coverage Are Too Narrow.
Only GPT-4 and Qwen3-32B are used, and GPT-4 is already an outdated baseline for an ACL 2026 paper. There are no 7B/8B small-model results, whereas ACON reports gains up to 46% for small models, suggesting different context-management behavior at smaller scales. If OCR-Memory's gains vanish on 8B because a smaller VLM cannot interpret dense renderings, its applicability shrinks substantially. This cannot currently be assessed.

⑧ [Engineering Reproducibility] No Public Code, Despite Highly Sensitive Rendering Details.
The paper specifies font (Arial 36pt), border width (3px), color (pure red), and interpolation (bicubic), precisely indicating that results are sensitive to rendering configuration. Glyph even uses LLM-driven genetic search to optimize rendering. Without code, third parties must blindly search the entire layout-hyperparameter space. The paper does not mention this among its limitations, yet it has the greatest practical impact.

These five papers are not a parallel list but a history of successively peeling away the same problem: first it does not fit; then it fits but selection is inaccurate; then abstraction loses detail; compression still loses information; finally, store the index in another space without discarding a single word of content.

2.1 MemoryBank — AAAI 2024

Problem addressed. LLM dialogue systems lacked cross-session long-term memory. MemoryBank established a full storage → retrieval → update cycle and introduced the Ebbinghaus forgetting curve to evolve memory strength with time and rehearsal frequency. In one sentence, it first made forgetting a first-class design element of LLM memory systems.

Unresolved problem. Its forgetting is irreversible deletion-based decay: once strength falls below a threshold, the record disappears permanently. Retrieval still uses pure semantic similarity—the “topically related yet logically irrelevant” behavior OCR-Memory criticizes. The opening is: can memory decay without being destroyed?

How the core paper responds. OCR-Memory rewrites forgetting from deletion into resolution degradation(Eq. 13), paired with active-recall upsampling (Eq. 14–15): memory blurs but never disappears, restoring original fidelity when retrieved. This is a direct adversarial rewrite of MemoryBank's forgetting. Quantitatively: Mind2Web Ele Acc 43.8 → 53.8 (+10.0pp / +22.8%), AppWorld Avg 52.1 → 58.1 (+6.0pp).

Transfer of key designs. MemoryBank's time-dependent strength function directly corresponds to OCR-Memory's aging function ρ(Δt): almost isomorphic forms, but the range changes from a [0,1] strength scalar to a discrete resolution level ℓ.

2.2 Agent Workflow Memory (AWM) — arXiv 2409.07429, 2024-09

Problem addressed. Web agents start from scratch each time. AWM induces reusable workflows from past trajectories and injects them into prompts as needed. Original abstract figures: Mind2Web +24.6% relative success, WebArena +51.1% relative success, fewer steps for successful tasks; online AWM exceeds baselines across task / website / domain settings by 8.9–14.0 absolute points. In one sentence, memory need not be raw experience; it can be procedural knowledge distilled from experience.

Unresolved problem. Abstraction is one-way and lossy. Once a workflow is induced, exact errors, intermediate states, and button wording cannot be recovered. The opening is: can abstraction's reusability coexist with raw-level traceability?

How the core paper responds. OCR-Memory's answer is Integration Rather Than Opposition: retain a high-level index—the visual map—but require verbatim text beneath it (Eq. 5's {s_{i,k}}). Locate-and-transcribe is a two-level structure: locate in abstraction, then fetch from the original. OCR-Memory beats AWM throughout Table 1 (Ele Acc 49.1→53.8, +9.6%; AW Avg 55.0→58.1, +5.6%), but the gap is smaller than against MemoryBank, indicating that abstraction works; OCR-Memory wins by preserving the raw foundation beneath it.

Transfer of key designs. AWM's selective workflow provision evolves into τ=0.4 ∪ TopK(5), capped at 20 paragraphs per query, a bounded injection policy that changes the question from which workflows to inject to which original passages to inject.

2.3 ACON — arXiv 2510.00615 (v3 2026-06), ICML 2026

Problem addressed. Long-horizon agents' contexts grow without bound. ACON uses natural-language optimization—analyzing failures and iteratively rewriting compression guidelines—to compress both observations and history, without fine-tuning. The abstract reports peak-token reductions versus existing compression baselines of 26–54% and gains on small models of up to +46% on AppWorld / OfficeBench / Multi-objective QA. In one sentence, compression guidelines themselves can be optimized, through natural language rather than gradients.

Unresolved problem. ACON is OCR-Memory's strongest representative of text-centric compression, whose fate is summarized as: “Text-centric compression inevitably trades off compression ratio against information fidelity.” ACON produces rewritten text; however good the guidelines, discarded words are gone. The opening is: is there a compression route outside text space?

How the core paper responds. This is the most direct transfer of the problem. OCR-Memory responds not by compressing better but by changing spaces: moving rate–distortion tradeoffs to pixels and using locate-and-transcribe to make distortion zero in the content dimension, leaving loss only in localization. ACON is runner-up at all three AppWorld difficulty levels (84.8 / 55.1 / 28.7), with OCR-Memory adding +1.4 / +2.3 / +2.1 pp. Yet ACON trails AWM on Mind2Web (48.2 vs. 49.1), suggesting general text compression struggles with highly structured observations such as web DOMs.

Transfer of key designs. ACON's iterative failure-driven compression optimization is replaced by a gradient-based version: weighted BCE (w₊=2.0 to favor recall) embodies the same failure-driven preference for including extra context rather than missing crucial evidence, moving from prompt-level natural-language optimization to parameters.

2.4 DeepSeek-OCR: Contexts Optical Compression — arXiv 2510.18234, 2025-10

Problem addressed. It proposed and experimentally validated a previously intuitive proposition: rendered text can be represented by far fewer visual tokens. Abstract figures: compression <10× gives 97% decoding accuracy, 20× gives about 60%; 100 visual tokens beat GOT-OCR2.0 at 256 tokens/page, and <800 tokens beat MinerU2.0's average 6000+ tokens; one A100-40G produces 200,000+ training pages daily. In one sentence, optical compression is not a metaphor: it has a measurable rate–distortion curve.

Unresolved problem. DeepSeek-OCR targets document parsing, where success means reading the words correctly, binding it to reconstruction—the steepest part of the curve, dropping toward 60% beyond 10×. The gap is: if the downstream task does not require reconstruction, does that curve still apply? Can redefining the task widen its usable range?

How the core paper responds. OCR-Memory's entire design answers this question: switching from reconstruction to localization changes the usable portion of the curve. It also yields a previously non-obvious implication: OCR-Memory empirically operates at 10.2–10.7× (Table 4), right at the edge of DeepSeek-OCR's 97% curve. This is no coincidence: the encoder's fidelity cliff fixes the operating point. It also explains why the paper avoids 20×, where DeepSeek-OCR falls to 60%.

Transfer of key designs. The most direct technical transfer: OCR-Memory's backbone are DeepSeek-OCR (3B), with the vision encoder fully frozen and LoRA only on the language decoder (r=16). OCR-Memory inherits how DeepSeek-OCR sees and replaces how it speaks—from text to 0/1 indices.

2.5 Set-of-Mark Prompting — arXiv 2310.11441, 2023-10

Problem addressed. GPT-4V understands images but struggles to specify which object it means. SoM's solution is strikingly simple: draw segmented regions directly on the image and label them with numbers so the model cites numbers rather than describes positions. In one sentence, discrete, identifiable symbolic anchors turn visual grounding from generation into selection.

Unresolved problem. SoM originally addresses spatial understanding of a single image. It does not answer whether the mechanism holds when images are rendered text rather than real scenes, marks grow from dozens to hundreds or thousands, and queries ask semantic relevance rather than spatial location.

How the core paper responds. OCR-Memory reinterprets SoM's spatial anchors as addresses of memory paragraphs(Eq. 7–8), with two upgrades: (1) output changes from free-text references to one binary logit per paragraph, enabling calibration (Eq. 11), thresholds, and TopK fallback (Eq. 12); (2) weighted BCE Training this identification ability instead of zero-shot prompting. Table 2 quantifies the transfer: bounding boxes without numbers lose 4.6 pp Ele Acc relative to full SoM, showing that discrete numbering as a symbolic anchor, rather than visual segmentation itself, drives the gain. This independently validates SoM's original conclusion in a new setting.

Transfer of key designs. The red-box-plus-number rendering convention is inherited almost unchanged: RGB 255,0,0 / 3px / Arial 36pt white on red. Only the marked objects change, from image objects to trajectory paragraphs.

2.6 Glyph (Parallel Work) — arXiv 2510.17800, 2025-10

Problem addressed. Another optical-compression approach appearing in the same month as DeepSeek-OCR, targeting not OCR but long-context modeling itself: render long text as images for VLM processing and use LLM-driven genetic search to optimize rendering. Abstract figures: 3–4× token compression, about 4× prefill/decode acceleration, about 2× SFT acceleration, at accuracy comparable to Qwen3-8B.

Glyph also leaves a problem OCR-Memory does not address: rendering configurations require search-based optimization. OCR-Memory fixes layout parameters manually without sensitivity analysis (§1.7-⑧); attaching Glyph's genetic search to its renderer is an almost ready-made improvement.

3.1 Evolution Tree (Nonlinear)

3.2 The Internal Logic of Evolution

The driving force can be summarized as a chain of progressively retreating failures:

Changes in the Nature of the Turning Points: I→II moves from human-written rules to evolving rules; II→III from rules to objective-driven optimization; and III→IV alone changes memory's physical representation. The first three ask how to cut text; the fourth first asks why it must be text.

4.1 Theory

(T1) Measure the Separation Between Localization and Reconstruction Information — High Value, Low Barrier.
Rather than broadly studying visual-memory theory, specifically: fix a corpus, vary rendering resolution, and jointly measure (a) retrieval Recall@k and (b) character-level OCR reconstruction accuracy; plot both curves. If (a) decays much more slowly than (b), this directly supports the central assumption (§1.7-⑥) and supplies a principled minimum resolution, replacing the handcrafted “latest 5 steps at high resolution.” With public DeepSeek-OCR weights and any long-document corpus, one A100 can finish this in a day.

(T2) Rate–Distortion Bounds for Optical Memory.
DeepSeek-OCR's 97%@10× / 60%@20× is an empirical curve for a particular encoder, not an information-theoretic bound. Given d×d pixels and a fixed font, what is the maximum achievable compression ratio for paragraph relevance classification? This determines whether 10× is near the limit or far below it.

(T3) Quantify Cross-Layer Tradeoffs.
OCR-Memory trades three resources: prompt tokens (−6.7×), disk (+81.7×), and latency (+5.7×). No work currently provides an equivalent exchange rate among them. An operational formalization: define Cost = α·tokens + β·bytes + γ·latency and sweep (α,β,γ) to identify the parameter-region boundary where OCR-Memory beats text-RAG. This directly answers which deployment scenarios suit optical memory.

4.2 Engineering and applications

(E1) Add the Missing Control: A Fine-Tuned Text Retriever.
As §1.7-③ explains, visual indexing's superiority over text indexing has not been fairly tested. Fine-tune a ~3B text reranker on the same HotpotQA distractor data for 3 epochs, then report Recall@1/5/10 and MRR under identical Mind2Web / AppWorld settings. The missing head-to-head needed to establish A over B is the most important next gap.

(E2) Break the Artificial 4096 Ceiling.
Run text-RAG and OCR-Memory at 4k / 8k / 16k / 32k / 128k budgets and locate the crossing. Below 16k, applicability narrows to edge / small-model deployment; above 128k, it becomes a general long-context solution. This experiment would be more informative than any existing table in the paper.

(E3) Automatically Search Rendering Configurations.
Attach Glyph's LLM-driven genetic search to OCR-Memory's renderer, searching font size, line spacing, column width, border width, numbering style, and number of resolution levels L. Current values are fixed manually without sensitivity analysis; Glyph suggests these parameters are not second-order effects.

(E4) Aging Functions in Dynamic Online Settings.
OCR-Memory's ρ(Δt) depends only on time, whereas MemoryBank's Ebbinghaus curve also considers rehearsal frequency. A direct improvement: ρ(Δt, n_recall)—slow aging according to n for a memory retrieved by active recall n times. Nearly free, since hits are already logged, this can alleviate §1.7-④'s self-fulfilling forgetting.

(E5) Natively Multimodal Trajectories.
Web-agent trajectories already contain screenshots. Current OCR-Memory renders text logs into images, but screenshots are images already. Naturally extend memory images to contain both rendered text and original screenshots under the same SoM index. This is an exclusive extension opportunity for optical versus text memory, untouched by the paper.

4.3 Emerging directions

(N1) Attach RL Memory Policies to Optical Memory — The First Meeting of Two Lines.
Memory-R1 / MemAct / AgeMem's policy internalization has not met OCR-Memory's modality reconstruction. Concretely, make paragraph grouping into images, upsampling timing, and τ selection the RL action space, rewarded by task success. All three decisions are currently handcrafted constants.

(N2) Generalize Reversible Compression.
Locate-and-transcribe is fundamentally lossy indexing plus lossless storage. Can it extend to KV cache(low-precision KV for attention routing, loading high-precision KV on a hit)? Or embedding(low-dimensional filtering, high-dimensional reranking—similar to two-stage ANN retrieval, but permanent promotion on a hit introduces new active-recall semantics)?

(N3) Shared Visual Memory and Security in Multi-Agent Systems.
Rendered images are readable across models: paragraphs indexed by a 3B VLM can be consumed by any text model, naturally supporting shared agent memory pools. But adversarial layouts—hidden text or misleading numbers—create a new memory-poisoning surface. No work currently studies optical-memory security.

(N4) Directly Attack Storage Costs.
The 81.7× disk expansion (Table 7) is the largest engineering weakness and can be optimized directly: current storage duplicates rendered images and complete text. If images reconstruct most text reliably—DeepSeek-OCR achieves <97% at 10×—could text copies be retained only for high-uncertainty paragraphs? This might bring disk expansion down to a single-digit factor.

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

• Proposes and empirically establishes experience-following: when task inputs closely match retrieved memory, agents produce highly similar outputs. This yields error accumulation with ineffective replay as failure modes and motivates using evaluation outcomes as memory-quality signals. The only selected candidate here that is purely empirical and proposes no new method, providing a failure-mode inventory for write policies.
• ACL | PDF | Code: not found

GAM: Hierarchical Graph-based Agentic Memory for LLM Agents (ACL 2026 Main, Long, pp. 34647–34664)

• Explicit Decouples memory encoding from consolidation, representing ongoing conversations in an event progression graph and integrating them into topic networks upon semantic drift. An interesting contrast: both use two levels, but GAM's levels are graph topology and OCR-Memory's are modalities.
• ACL | PDF | Code: not found

MAGMA: A Multi-Graph based Agentic Memory Architecture for AI Agents (ACL 2026 Main, Long, pp. 36848–36865)

• Criticizes semantic-similarity retrieval over monolithic stores for entangling temporal, causal, and entity information, replacing it with four separate semantic / temporal / causal / entity graphs. This and OCR-Memory's “topically related yet logically irrelevant” critique are two remedies for the same diagnosis—one separates graphs, the other changes modality. Worth reading together.
• ACL | PDF | Code: not found

Adaptive Memory Distillation for LLM Agents (ACL 2026 Main, Long)

• ACL 2026 main-track memory-distillation work, extending AWM's abstraction line into 2026. Paired with OCR-Memory, it offers the latest abstraction-versus-no-abstraction comparison.
• PDF | Code: not found

H-MEM: Hierarchical Memory for High-Efficiency Long-Term Reasoning in LLM Agents (EACL 2026, Long)

• Efficiency-oriented hierarchical memory. Compare its semantic-abstraction hierarchy directly with OCR-Memory's hierarchy of information fidelity.
• ACL | PDF

RecMem: Recurrence-based Memory Consolidation (Findings of ACL 2026)

• Uses recurrence for consolidation, along the parametric-versus-external-memory line. OCR-Memory is extremely externalized, with even the content on disk; RecMem moves in the opposite direction and serves as the other endpoint.
• PDF

Amory: Building Coherent Narrative-Driven Agent Memory (EACL 2026, Long)

• Narrative-coherence-driven organization focuses on relationships among records rather than individual representations, complementing this issue's theme.
• PDF

Governed Shared Memory for Multi-Agent LLM Systems (arXiv 2606.24535, 2026-06)

• A governance framework for multi-agent shared memory, directly related to §4.3-N3, but focused on text memory; optical-memory poisoning remains unstudied.
• arXiv

From Recall to Forgetting: Benchmarking Long-Term Memory for Personalized Agents (arXiv 2604.20006, 2026-04)

• A benchmark making forgetting itself the evaluation target. OCR-Memory's reversible forgetting needs exactly such a testbed; currently it is verified only indirectly on Mind2Web / AppWorld.
• arXiv

1. Exact Figure 2 values cannot be cited: the token-budget sweep is plotted without a table, so §1.5.8's transition analysis only describes qualitative trends.
2. OCR-Memory has no public code, preventing implementation cross-checking. The analytic form of the aging function ρ is actually not specified; it is only described as monotonic.
3. tabzhangjx.github.io/year-archive/ Fetch failed: deduplication used the explicit list in the task prompt rather than the online archive.
4. The paper does not state the exact Mind2Web Cross-Task test-set size: §1.7-①'s inference that 0.5 pp is a single-digit task difference uses the public understanding that this split has hundreds of tasks. If the actual size is much larger, that critique weakens.
5. AWM's formal conference publication could not be confirmed: only its arXiv page was found, not proceedings, so it is labeled a preprint.
6. No hyperparameter sensitivity sweeps are reported(τ, resolution levels L, high-resolution step count). This report does not invent curves; §4 lists these as missing experiments.