EchoPath: Execution-Level Replayable Memory for GUI Agents
Yao Zhao, Aditya Shanmugham, Swastik Roy, Yanxun Xu
cs.AI
Quick verdict
GUI-trajectory memories replayed via image re-aiming cut tokens >90% and time ~60% on OSWorld reruns; solid systems work but only same-task reuse is tested.
Summary
EchoPath converts artifact-validated GUI agent trajectories into structured, MCP-style callable memories that a host agent invokes for recurrent desktop tasks instead of re-running a fresh observe-plan-ground-act loop. Each memory stores retrieval keys, application/state preconditions, typed flexible parameters, visual evidence, validation provenance, and lifecycle state. Replay avoids coordinate copying via an image-based target-reaiming (IBTR) algorithm: stored interface crops are template-matched against the current screenshot (correlation score with ambiguity margin, 0.5–2.0 scale sweep) to emit corrected pointer coordinates, with bounded step-local grounding repair or fresh planning as fallbacks. In a paired two-pass experiment on OSWorld-Verified (second pass at 1600×900 to forbid coordinate copying), replay across Codex, Claude Code, and Kimi Code reached 87.3–92.8% success with medians of 20–30k tokens and ~130 s, versus 586k tokens and 316 s for a Synapse trajectory-prompt baseline; the first pass cost a median ~572k tokens. Component diagnostics report 100% retrieval recall up to 659 memories, sub-2-pixel re-aiming error under random scaling, <1 MB storage per memory, and 100% flexible-input enforcement offline. Main limitations: the second pass reruns the same tasks (parameter rebinding tested only offline on 50 memories with synthetic reasoner responses), a single baseline on one model, no significance intervals, only 78% rejection accuracy for incompatible start states, and EchoPath/OpenPath plus IBTR/IMTR naming inconsistencies.
Review complete
Assessment at a glance
5
risk flags identified
● high 0
● medium 4
● low 1
4 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- Artifact-validated GUI trajectories can be standardized into MCP-style callable memories and replayed deterministically when compatibility gates pass (§1–2).
- The IBTR image-matching algorithm re-aims stored pointer actions against the current screen, bypassing per-step LLM grounding, and remains accurate under resolution change (§2.2, Table 2).
- Second-pass replay matches fresh-planning success (87.3–92.8% vs 91.8% Synapse) at roughly 3–7% of the tokens and ~40% of the execution time (Table 1).
- Component diagnostics show 100% retrieval recall to 659 memories, sub-2-pixel re-aiming under scaling, <1 MB per memory, and 100% flexible-input enforcement offline (§4, Tables 2–3, Fig. 4).
Strengths
Anti-shortcut experimental design
Second pass deliberately shifts resolution 1920×1080→1600×900 (§3.1) so coordinate copying fails by construction, and the offline IBTR test adds a random 0.6–1.6× scaling condition — replay success must come from genuine visual re-aiming.
Fully specified deterministic core
Supp. B.3 + Algorithm 1 publish all thresholds (η=0.78, ηmargin=0.02, scale 0.5–2.0 step 0.05, σflat=1.0), the correlation/flat-crop formulas, ambiguity-margin rejection, and frame-conversion math, making the binding operator ρ reproducible without reading code.
Model-agnostic claim actually tested across harnesses
Unusual pipeline transparency
Weaknesses
Same-task reruns carry the headline numbers
Thin baseline coverage
No uncertainty quantification
Safety margin of the state gate
Artifact and naming inconsistencies
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
EchoPath (arXiv 2609.16635, cs.AI; Johns Hopkins + Amazon AGI) addresses recurrent GUI work: rather than re-running an observe-plan-ground-act loop for every repeated task, it converts artifact-validated agent trajectories into standardized, MCP-style callable memories gated by preconditions and lifecycle state. Four claims: (1) validated trajectories can be packaged as deterministic callable procedures; (2) the image-based target-reaiming (IBTR) algorithm binds stored pointer actions to the current screen, bypassing per-step LLM grounding; (3) a model-agnostic harness cuts median second-pass token cost by >90% and time by ~60% on OSWorld-Verified; (4) memories carry governed lifecycles (promotion, branching, quarantine). The authors scope the claim carefully to recurrence under compatible state — not general transfer.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 4 strengths & 5 weaknesses, analyzed in detail
- 5 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is EchoPath?
EchoPath is a harness that turns successful GUI agent trajectories into structured, reusable 'callable memories' similar to MCP tool calls. On repeat desktop tasks, the agent invokes a stored procedure directly instead of re-planning each step; stored coordinates are treated as visual evidence and re-aimed against the current screen by image matching, with automatic fallback to normal agent planning when replay is unsafe.
How much cheaper is EchoPath replay than fresh agent execution?
On OSWorld-Verified tasks rerun a second time, replay cost a median of about 20,000–30,000 tokens and roughly 130 seconds across Codex, Claude Code, and Kimi Code clients, versus 586,386 median tokens and 316 seconds for the Synapse trajectory-prompt baseline; the first fresh pass cost a median of about 572,000 tokens and 4.5 minutes. Success rates were comparable (87.3–92.8%).
Does EchoPath still work when the screen resolution or layout changes?
Resolution changes are handled by design: the evaluation lowered resolution from 1920×1080 to 1600×900 in the second pass, and an offline test under random 0.6–1.6× scaling kept re-aiming error within 2 pixels for 188 of 200 cases. However, the authors acknowledge remaining vulnerability to toolbar rearrangements, localization, responsive layouts, and near-duplicate interface elements; live drift robustness was not tested.
Is EchoPath's code and data available?
The paper links a public GitHub repository (github.com/JackZhao1998/EchoPath) and publishes all pipeline prompts in its supplement. Experiments ran on public OSWorld-Verified tasks using CPU-only AWS t3.large VMs and commercial agent CLIs (Codex, Claude Code, Kimi Code), so reproducing headline numbers requires those paid APIs plus OSWorld VM setup.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Text extraction elided several spans (parts of §2.1, the tail of Table 2, §B.2 gate details, later references), so a few numbers are taken from partially truncated passages.
- Figures 1–4 are referenced but not analyzable from the extracted text; figure-dependent claims (e.g., construction-cost inset, storage curve) rest on in-text descriptions only.
- The GitHub repository's actual contents and the identity of the released package (OpenPath vs EchoPath) could not be verified without network access.
- Supplementary Section D prompts are truncated at the end (D.7 partial), so fallback-prompt fidelity is assumed rather than verified.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.