{
  "api_version": "v1",
  "id": "effective-context-memory-2026-10-29",
  "published_at": "2026-10-29",
  "publisher": "Darkoza Knowledge Lab",
  "title": "Effective context capacity: external memory and bounded evidence packs",
  "evidence_status": "reproducible synthetic engineering benchmark; standalone research prototype",
  "production_status": "not connected to the public OERSI inference route",
  "method": {
    "storage": "SQLite; original document content retained with SHA-256 digest",
    "retrieval": "BM25-style lexical ranking with a small explicit-priority factor",
    "context_pack": "whole retrieved chunks with source labels and document/chunk hashes",
    "token_estimator": "ceil(UTF-8 bytes / 4); heuristic, not an exact model tokenizer",
    "data_scope": "synthetic only",
    "model_calls": 0,
    "network_calls": 0
  },
  "metrics": {
    "synthetic_records": 1000,
    "queries": 200,
    "baseline": {
      "name": "recency-only context window",
      "last_records_kept": 24,
      "hits": 6,
      "hit_rate": 0.03
    },
    "candidate": {
      "name": "external lexical retrieval and bounded context pack",
      "retrieval_top_5_hits": 200,
      "retrieval_top_5_hit_rate": 1.0,
      "packed_fact_hits": 200,
      "packed_fact_hit_rate": 1.0,
      "estimated_context_budget": 256,
      "max_observed_estimated_context": 173
    },
    "latency_ms": {
      "run_1": {
        "median_search": 23.9316,
        "p95_search": 34.5329
      },
      "run_2": {
        "median_search": 23.4325,
        "p95_search": 25.3504
      },
      "median_search_range_ms": [
        23.4325,
        23.9316
      ],
      "p95_search_range_ms": [
        25.3504,
        34.5329
      ],
      "scope": "two runs on one server with a synthetic corpus; not a production SLO"
    },
    "integrity": {
      "documents_checked": 1000,
      "chunks_checked": 1000,
      "hash_mismatches": 0
    },
    "unit_tests_passed": 8,
    "unit_tests_total": 8,
    "benchmark_runs": 2,
    "deterministic_results_match_across_runs": true
  },
  "interpretation": "The benchmark verifies retrieval mechanics on synthetic data with target identifiers present in each query. It does not evaluate semantic paraphrase, LLM answer correctness, long-horizon reasoning, real-world corpora or production performance. The 100% hit rate is not a general memory or reasoning score.",
  "promotion_gate": {
    "status": "proposed; not yet tested on public long-memory QA",
    "minimum_held_out_answer_accuracy_gain_percentage_points": 10,
    "requirements": [
      "same model and input/output token budget for candidate and baseline",
      "held-out public benchmark data used within its license terms",
      "measure unsupported-claim rate and citation precision",
      "zero unauthorized consequential actions",
      "tenant isolation, retention, deletion and backup/restore checks"
    ]
  },
  "artifacts": {
    "report_html": "https://darkoza.com/science/reports/effective-context-memory/report.html",
    "report_markdown": "https://darkoza.com/science/reports/effective-context-memory/report.md",
    "raw_results": "https://darkoza.com/science/reports/effective-context-memory/results.json",
    "implementation": "https://darkoza.com/science/reports/effective-context-memory/context_memory.py",
    "benchmark_source": "https://darkoza.com/science/reports/effective-context-memory/benchmark.py",
    "unit_tests": "https://darkoza.com/science/reports/effective-context-memory/test_context_memory.py",
    "repeat_results": "https://darkoza.com/science/reports/effective-context-memory/results-repeat.json"
  },
  "references": [
    "https://papers.neurips.cc/paper_files/paper/2025/hash/a4ca07aa108036f80cbb5b82285fd4b1-Abstract-Conference.html",
    "https://proceedings.iclr.cc/paper_files/paper/2026/hash/af1fd844255b6bc846efa1bec0eb6411-Abstract-Conference.html",
    "https://arxiv.org/abs/2310.08560",
    "https://arxiv.org/abs/2606.09079",
    "https://hai.stanford.edu/ai-index/2026-ai-index-report/technical-performance"
  ]
}