{ "schema_version": 1, "builder_plan": { "scenario": "A support or data assistant repeats work across related customer requests.", "change": "Try the paper's memory change in a replay of past tasks, with separate memory per customer.", "baseline": "Your current memory pipeline plus a no-memory control using the same agent.", "steps": [ "Locate the paper's implementation and check its license, model support, data needs, and hardware requirements. Stop here if you cannot obtain the required artifacts.", "Save a small set of representative product tasks and expected outcomes. Keep evaluation tasks out of any training or tuning, and choose quality and cost limits before running.", "Run the baseline and the proposed change on identical inputs. Keep other settings fixed and repeat paired runs to expose variation.", "Save per-task outputs, failures, timings, and all usage costs. Review regressions before expanding the test." ], "measure": "Task success, stale memories, retrieval misses, response latency, and all memory maintenance costs.", "decision": "Keep the change only if it meets your preselected quality and latency limits and improves the metric you care about after setup and operating costs. Otherwise retain the baseline. This is an exploration plan, not a validated recommendation." }, "builder_review": null, "arxiv_id": "2610.20927", "title": "Prune the Summary, Not the History: Source-Judged Compaction for Long-Horizon Agents", "summary": "SJC replaces nothing architecturally: it is a training-free post-processing step layered on an existing compaction pipeline, scoring each summary entry for provenance in environment outputs or the prior summary and deleting unsourced entries, leaving the agent and compressor unchanged. Its cost is one additional inexpensive judge request per compaction event; the judge is a non-generative calibrated decision model, so no generative inference, fine-tuning, or extra context tokens are required. It can fail if the judge miscalibrates and prunes genuinely useful reasoning entries, and the reported gains are validated only on AppWorld and OfficeBench, so transfer to other agent workloads is unproven.", "rationale": "The method is cheap, training-free, and compatible with API-based or single-GPU deployments, but its 2.1%-8.9% reliability gain should be validated on the reader's own compaction pipeline and workloads before production adoption.", "technique": "Source-Judged Compaction (summary entry pruning)", "family": "memoria_e_contexto", "recommendation": "testar", "published": "2026-10-14", "as_of": "2026-10-18", "editorial_status": "indexed", "source_url": "https://arxiv.org/abs/2610.20927", "independent_implementations": 1, "total_implementations": 2, "stars": 132, "citations": 0, "claims": [ { "claim_id": "claim-01", "statement": "SJC replaces nothing architecturally: it is a training-free post-processing step layered on an existing compaction pipeline, scoring each summary entry for provenance in environment outputs or the prior summary and deleting unsourced entries, leaving the agent and compressor unchanged. Its cost is one additional inexpensive judge request per compaction event; the judge is a non-generative calibrated decision model, so no generative inference, fine-tuning, or extra context tokens are required. It can fail if the judge miscalibrates and prunes genuinely useful reasoning entries, and the reported gains are validated only on AppWorld and OfficeBench, so transfer to other agent workloads is unproven.", "basis": "inferred", "result": "", "baseline": "", "conditions": "", "source_url": "https://arxiv.org/abs/2610.20927", "source_page": null, "source_excerpt": "" }, { "claim_id": "claim-02", "statement": "SJC improves cross-run reliability over the underlying compressor by 2.1%-8.9% on AppWorld and OfficeBench with two agent models, without increasing agent token usage; in the paper's analysis, 83% of failed runs traced to unsupported summary entries.", "basis": "inferred", "result": "", "baseline": "", "conditions": "", "source_url": "https://arxiv.org/abs/2610.20927", "source_page": null, "source_excerpt": "" } ], "exposure_map": [ { "dimension": "quality", "basis": "not_evaluated", "finding": "", "interpretation": "", "limitation": "", "next_check": "", "source_url": "", "source_page": null, "source_excerpt": "" }, { "dimension": "compute", "basis": "not_evaluated", "finding": "", "interpretation": "", "limitation": "", "next_check": "", "source_url": "", "source_page": null, "source_excerpt": "" }, { "dimension": "latency", "basis": "not_evaluated", "finding": "", "interpretation": "", "limitation": "", "next_check": "", "source_url": "", "source_page": null, "source_excerpt": "" }, { "dimension": "operations", "basis": "not_evaluated", "finding": "", "interpretation": "", "limitation": "", "next_check": "", "source_url": "", "source_page": null, "source_excerpt": "" }, { "dimension": "compatibility", "basis": "not_evaluated", "finding": "", "interpretation": "", "limitation": "", "next_check": "", "source_url": "", "source_page": null, "source_excerpt": "" }, { "dimension": "security", "basis": "not_evaluated", "finding": "", "interpretation": "", "limitation": "", "next_check": "", "source_url": "", "source_page": null, "source_excerpt": "" }, { "dimension": "data_and_training", "basis": "not_evaluated", "finding": "", "interpretation": "", "limitation": "", "next_check": "", "source_url": "", "source_page": null, "source_excerpt": "" }, { "dimension": "reproducibility", "basis": "not_evaluated", "finding": "", "interpretation": "", "limitation": "", "next_check": "", "source_url": "", "source_page": null, "source_excerpt": "" } ], "risks": [], "minimum_test": [], "open_questions": [], "independent_tests": [], "lab_reproductions": [], "report_available": false, "validation_tier": "unknown", "evidence_tier": "unknown", "infrastructure_basis": "unknown", "equations": [], "equations_status": "not_fetched", "core_kind": "", "equations_fetched_at": "", "equations_source_url": "", "equations_source_sha256": "", "exposures_reviewed_at": "", "exposures_source_sha256": "" }