{ "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": "2609.17346", "title": "Where Should a Document Live: Context, Representations, or Parameters?", "summary": "Cartridges replace passing documents in the context window (or fine-tuning them into parameters) with pre-trained KV-cache representations that are loaded at inference time. The cost is a training pass per document collection to produce the cartridges, storage of the resulting KV caches, and a 6% degradation on control benchmarks (13% on coding) due to catastrophic forgetting. The main failure modes are forgetting on unrelated tasks and degraded accuracy when Compaction is substituted at compression rates above 50x, where it trails parametric methods by 10 points.", "rationale": "The paper shows in-context retrieval already matches the best injection method, so a team on one 24 GB GPU gains nothing by adopting a per-corpus training pipeline that also introduces measurable forgetting.", "technique": "Cartridges: trained KV-cache knowledge injection", "family": "memoria_e_contexto", "recommendation": "observar", "published": "2026-09-15", "as_of": "2026-10-18", "editorial_status": "indexed", "source_url": "https://arxiv.org/abs/2609.17346", "independent_implementations": 0, "total_implementations": 1, "stars": 0, "citations": 0, "claims": [ { "claim_id": "claim-01", "statement": "Cartridges replace passing documents in the context window (or fine-tuning them into parameters) with pre-trained KV-cache representations that are loaded at inference time. The cost is a training pass per document collection to produce the cartridges, storage of the resulting KV caches, and a 6% degradation on control benchmarks (13% on coding) due to catastrophic forgetting. The main failure modes are forgetting on unrelated tasks and degraded accuracy when Compaction is substituted at compression rates above 50x, where it trails parametric methods by 10 points.", "basis": "inferred", "result": "", "baseline": "", "conditions": "", "source_url": "https://arxiv.org/abs/2609.17346", "source_page": null, "source_excerpt": "" }, { "claim_id": "claim-02", "statement": "In the realistic multi-document retrieval setting, Cartridges are the only injection method matching in-context learning, leading parametric methods by 29 points and Compaction by 15 points; in the oracle setting they beat parametric methods by about 10 points at nearly every storage budget.", "basis": "inferred", "result": "", "baseline": "", "conditions": "", "source_url": "https://arxiv.org/abs/2609.17346", "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": "" }