REPRODUCIBILITY.md

Reproducibility

This document records the current reproducibility path for the active root manuscript, main.tex. The repository is intentionally split into deterministic non-API experiments, cached external LongMemEval artifacts, and API reader runs.

Environment

Use Python 3.10 or newer.

python -m pip install -r requirements.txt

Optional dependencies are separated by task:

python -m pip install -r requirements-api.txt
python -m pip install -r requirements-milp.txt

The exact-small MemAudit benchmark and unit tests use only the Python standard library plus pytest for tests. LongMemEval retrieval regeneration uses local ML dependencies and downloads the LongMemEval-S dataset and dense retriever model. API reader runs use OpenRouter and require an API key.

LaTeX Build

On this machine, latexmk, pdflatex, and tectonic were not available on PATH during the 2026-04-28 local check. The attempted local build is recorded in latex_compile_attempt.txt. A generated latex_compile.log also exists locally, but *.log is ignored by the repository.

If a TeX distribution is installed locally, run one of:

make paper
make paper-pdflatex
make paper-tectonic

Because local compilation was unavailable here, .github/workflows/latex.yml builds main.tex with GitHub Actions on push and pull request.

Unit Tests

python -m pytest test_oraclemem.py

Current verification on 2026-05-01: both python -m unittest test_oraclemem.py and python -m pytest test_oraclemem.py ran 17 tests and passed.

Quickcheck

Use this before any expensive API or GPU work:

python -m unittest test_oraclemem.py
python run_oraclemem_mvp.py --n-seeds 3 --budgets 4 --distribution base --methods opt,oracle_gvt,density_only --out-dir oraclemem_runs/quickcheck

Expected outputs:

• oraclemem_runs/quickcheck/raw_results.jsonl
• oraclemem_runs/quickcheck/summary.json
• oraclemem_runs/quickcheck/summary.md

Exact-Small Benchmark

Used by the exact-small budget-sweep figure in main.tex.

python run_oraclemem_mvp.py \
  --n-seeds 500 \
  --budgets 0.01,0.02,0.05,0.10,0.20 \
  --distribution base \
  --methods opt,oracle_gvt,density_only,recency_raw,summary_only,fact_only,no_tombstone_gvt,no_tombstone_opt \
  --out oraclemem_runs/exact_500

Expected outputs:

• oraclemem_runs/exact_500/raw_results.jsonl
• oraclemem_runs/exact_500/summary.json
• oraclemem_runs/exact_500/summary.md

The reported ratio_to_opt field is valid only for these exact-small runs where the denominator is an exact certified optimum.

Stress Suite

Used by the validity-heavy stress figure in main.tex. The manuscript reports the validity-heavy subset base, update_chain, and temporal_interval from the larger stress artifact.

python run_oraclemem_mvp.py \
  --n-seeds 500 \
  --budgets 0.02,0.05,0.10,0.20 \
  --distribution base,update_chain,temporal_interval,density_trap,scope_shift,summary_tradeoff,redundancy_heavy,abstention_hard \
  --methods opt,oracle_gvt,density_only,greedy,recency_raw,reservoir_raw,summary_only,fact_only,no_tombstone_gvt,no_tombstone_opt \
  --out oraclemem_runs/stress_exact_500

Expected outputs:

• oraclemem_runs/stress_exact_500/raw_results.jsonl
• oraclemem_runs/stress_exact_500/summary.json
• oraclemem_runs/stress_exact_500/summary.md

Representative Non-Oracle Writers

Used by the text diagnostic on Estimated-GVT, A-MAC-like admission, and Mem0-style extraction proxies. These methods use visible candidate features, not hidden coverage labels.

python run_oraclemem_mvp.py \
  --n-seeds 500 \
  --budgets 4,6 \
  --distribution base,update_chain,temporal_interval \
  --methods opt,oracle_gvt,estimated_gvt,amac_admission,mem0_extract,density_only,recency_raw,summary_only,fact_only,no_tombstone_opt \
  --out-dir oraclemem_runs/representative_writers_500

Expected outputs:

• oraclemem_runs/representative_writers_500/raw_results.jsonl
• oraclemem_runs/representative_writers_500/summary.json
• oraclemem_runs/representative_writers_500/summary.md

No-API Proxy Writer Baselines

Diagnostic only; not a main-paper result after the 9-page compression pass. This local diagnostic addresses the real-system-comparison concern without calling OpenRouter, OpenAI, embedding services, or API reader code. It runs deterministic proxies for MemGPT-style tiering, Mem0-style extraction, A-Mem-style graph/evolving memory, and A-MAC-style admission under the same MemAudit candidate protocol and exact OPT denominator.

python run_oraclemem_mvp.py \
  --n-seeds 50 \
  --distribution base,update_chain,scope_shift_v2,density_trap_v2,temporal_interval \
  --budgets 4,6 \
  --methods opt,oracle_gvt,memgpt_tiered,mem0_extract,amem_graph,amac_admission,generic_candidate_opt,no_tombstone_opt \
  --out-dir oraclemem_runs/proxy_writer_baselines_50 \
  --enable-retrieval \
  --retrieval fixed,oracle

Expected outputs:

• oraclemem_runs/proxy_writer_baselines_50/raw_results.jsonl
• oraclemem_runs/proxy_writer_baselines_50/summary.json
• oraclemem_runs/proxy_writer_baselines_50/summary.md
• oraclemem_runs/proxy_writer_baselines_50/REPORT.md

The report is explicit that these local ratios are synthetic exact-small ratios for proxy writers. A real-system comparison still requires running the actual systems with budget-matched memory generation, storage accounting, retrieval configuration, and evaluation traces.

Gemini Natural Coverage Pilot

Superseded by the Natural-200 and adjudicated-subset results in main.tex. This run builds a smaller LongMemEval-S support-slice MemAudit coverage package using Gemini through OpenRouter. It requires api.env with OPENROUTER_API_KEY. Candidate generation receives only support sessions plus distractors; query/gold-answer fields are used only in the separate labeling step.

python llm_memory_validation/gemini_natural_oraclemem.py \
  --limit 50 \
  --distractors-per-example 2 \
  --budgets 30,60,100 \
  --out-dir llm_memory_validation/gemini_natural_oraclemem_50 \
  --request-sleep 0.02

python scripts/audit_coverage_artifacts.py \
  --no-defaults \
  --artifact gemini_natural_50=llm_memory_validation/gemini_natural_oraclemem_50/coverage_package \
  --output-dir llm_memory_validation/gemini_natural_oraclemem_50/coverage_audit

Expected outputs:

• llm_memory_validation/gemini_natural_oraclemem_50/REPORT.md
• llm_memory_validation/gemini_natural_oraclemem_50/coverage_resolved_summary.json
• llm_memory_validation/gemini_natural_oraclemem_50/coverage_package/
• llm_memory_validation/gemini_natural_oraclemem_50/coverage_audit/REPORT.md

The first uncached 50-example run used 248 API calls, 502,698 total tokens, and about $0.286 in OpenRouter-reported cost. Cached reruns use zero additional API calls. This run is a pilot: 30/50 examples are coverage-resolved and the labels are single-model annotations rather than human adjudications.

Natural-200 And Model-Adjudicated Subsets

Used by the natural package reliability table and the model-adjudicated subset table in main.tex.

Primary Natural-200 package:

python llm_memory_validation/gemini_natural_oraclemem.py \
  --limit 200 \
  --distractors-per-example 0 \
  --max-session-words 1800 \
  --budgets 30,60,100 \
  --out-dir llm_memory_validation/oraclemem_natural_200_gemini_v2 \
  --request-sleep 0.02

python scripts/audit_coverage_artifacts.py \
  --no-defaults \
  --artifact natural_200_gemini_v2=llm_memory_validation/oraclemem_natural_200_gemini_v2/coverage_package \
  --output-dir llm_memory_validation/oraclemem_natural_200_gemini_v2/coverage_audit

Gemini Flash adjudicated subset:

python llm_memory_validation/adjudicate_natural_package.py \
  --primary-package-dir llm_memory_validation/oraclemem_natural_200_gemini_v2/coverage_package \
  --out-dir llm_memory_validation/natural_adjudicated_100_gemini_flash \
  --model google/gemini-2.5-flash \
  --limit 100 \
  --budgets 30,60,100 \
  --secondary-agreement-rows llm_memory_validation/natural50_annotation_agreement_gemini31_vs_gemini25/agreement_rows.jsonl \
  --mem0-raw-results llm_memory_validation/mem0_natural200_actual/raw_results.jsonl \
  --request-sleep 0.02

Gemini 3.1 Flash-Lite spot-check:

python llm_memory_validation/adjudicate_natural_package.py \
  --primary-package-dir llm_memory_validation/oraclemem_natural_200_gemini_v2/coverage_package \
  --out-dir llm_memory_validation/natural_spotcheck_30_gemini31_flash_lite \
  --model google/gemini-3.1-flash-lite-preview \
  --limit 30 \
  --budgets 30,60,100 \
  --methods opt,oracle_gvt,estimated_gvt,amac_admission,summary_only,fact_only,recency_raw \
  --secondary-agreement-rows llm_memory_validation/natural50_annotation_agreement_gemini31_vs_gemini25/agreement_rows.jsonl \
  --mem0-raw-results llm_memory_validation/mem0_natural200_actual/raw_results.jsonl \
  --request-sleep 0.02 \
  --skip-existing

python scripts/audit_coverage_artifacts.py \
  --no-defaults \
  --artifact natural_spotcheck_30_gemini31_flash_lite=llm_memory_validation/natural_spotcheck_30_gemini31_flash_lite/coverage_package \
  --output-dir llm_memory_validation/natural_spotcheck_30_gemini31_flash_lite/coverage_audit

The Flash-Lite spot-check attempted 30 examples, exported 29 accepted/corrected examples, rejected 1, used 201,301 total tokens, and cost $0.0639 through OpenRouter. It is model adjudication, not human validation.

Human-Edited Natural Seed Package

This package is a fictional 100-example natural-memory seed set that was manually edited/audited after generation. It is used as an artifact-validity check for manual annotation plus exact finite-package scoring. It is not an inter-annotator agreement study.

Validate the canonical JSONL:

python scripts/validate_human_style_examples.py llm_memory_validation/human_style_examples/examples_100.jsonl

Evaluate the finite package with an exact dynamic-programming denominator:

python llm_memory_validation/evaluate_human_style_examples.py \
  --examples-jsonl llm_memory_validation/human_style_examples/examples_100.jsonl \
  --out-dir llm_memory_validation/human_style_examples/eval_package_100 \
  --budgets 150,300,600,1000 \
  --methods opt,oracle_gvt,estimated_gvt,memgpt_tiered,amem_graph,amac_admission,mem0_extract,density_only,greedy,fact_only,summary_only,recency_raw,no_tombstone_opt

Expected outputs:

• llm_memory_validation/human_style_examples/eval_package_100/raw_results.jsonl
• llm_memory_validation/human_style_examples/eval_package_100/summary.json
• llm_memory_validation/human_style_examples/eval_package_100/summary.md
• llm_memory_validation/human_style_examples/eval_package_100/REPORT.md

Current verification on 2026-05-01: validation passed with 100 records and no structural errors. The evaluator reports the denominator as exact_human_audited_package_dp.

Export the same examples to the shared coverage-package schema:

python llm_memory_validation/export_human_style_coverage_package.py \
  --examples-jsonl llm_memory_validation/human_style_examples/examples_100.jsonl \
  --out-dir llm_memory_validation/human_style_examples/coverage_package

python scripts/audit_coverage_artifacts.py \
  --no-defaults \
  --artifact human_style_coverage=llm_memory_validation/human_style_examples/coverage_package \
  --output-dir llm_memory_validation/human_style_examples/coverage_package_audit

Run actual public A-Mem on the exported human-edited package:

python llm_memory_validation/run_actual_amem_natural_baseline.py \
  --package-dir llm_memory_validation/human_style_examples/coverage_package \
  --out-dir llm_memory_validation/human_style_examples/actual_amem_gemini_flash_100 \
  --limit 100 \
  --budgets 150,300,600,1000,5000 \
  --amem-model google/gemini-2.5-flash \
  --coverage-model google/gemini-2.5-flash \
  --request-sleep 0.02 \
  --amem-max-tokens 3000

Current actual A-Mem human-edited run: 85 query-resolved examples, 456 cached API prompts, 269,742 tokens, estimated OpenRouter cost $0.233. Full A-Mem notes reach union-OPT ratio 0.971 at all reported budgets; metadata-only reaches 0.247. This result is strong but should be interpreted with the package caveat: the sessions are short enough that full notes fit the 150+ word budgets.

Learned Writer Transfer Diagnostic

This local run trains a visible-feature utility estimator on train-only oracle labels from synthetic instances plus the Natural-200 model-annotated package, then evaluates held-out decisions on the human-edited seed package. Hidden coverage is used for train labels only; held-out selection sees visible candidate metadata only.

python llm_memory_validation/evaluate_learned_writer_transfer.py \
  --out-dir llm_memory_validation/human_style_examples/learned_writer_transfer \
  --budgets 150,300,600,1000 \
  --methods opt,oracle_gvt,estimated_gvt,estimated_utility,memgpt_tiered,amem_graph,amac_admission,mem0_extract,density_only,greedy,fact_only,summary_only,recency_raw,no_tombstone_opt

Expected outputs:

• llm_memory_validation/human_style_examples/learned_writer_transfer/raw_results.jsonl
• llm_memory_validation/human_style_examples/learned_writer_transfer/summary.json
• llm_memory_validation/human_style_examples/learned_writer_transfer/summary.md
• llm_memory_validation/human_style_examples/learned_writer_transfer/REPORT.md
• llm_memory_validation/human_style_examples/learned_writer_transfer/train_manifest.json

Current run: 1,000 synthetic train instances plus 200 natural train instances with 22,106 train candidates. Estimated-GVT reaches held-out exact package-OPT ratios 0.933/0.926/0.854/0.792 at budgets 150/300/600/1000. This is a deployable-writer diagnostic, not an inter-annotator natural benchmark.

Training-source ablations:

python llm_memory_validation/evaluate_learned_writer_transfer.py \
  --out-dir llm_memory_validation/human_style_examples/learned_writer_transfer_synth_only \
  --train-natural-limit 0 \
  --budgets 150,300,600,1000 \
  --methods opt,oracle_gvt,estimated_gvt,estimated_utility,amac_admission,mem0_extract,density_only,greedy,fact_only,summary_only,recency_raw,no_tombstone_opt

python llm_memory_validation/evaluate_learned_writer_transfer.py \
  --out-dir llm_memory_validation/human_style_examples/learned_writer_transfer_natural_only \
  --n-synthetic-train-seeds 0 \
  --budgets 150,300,600,1000 \
  --methods opt,oracle_gvt,estimated_gvt,estimated_utility,amac_admission,mem0_extract,density_only,greedy,no_tombstone_opt

Current ablations: synthetic-only Estimated-GVT reaches 0.667/0.778/0.792/0.833; Natural-200-only reaches 0.000/0.074/0.375/0.486. The combined run is therefore the paper-facing learned-writer result because it is strongest at tight and medium budgets.

Natural Writer Adapter Diagnostic

This local run scores Letta/MemGPT-style archival/recency and A-Mem-style graph adapters on the adjudicated natural package under the same exact package OPT denominator. It does not call an API and does not run Letta or A-Mem itself.

python llm_memory_validation/evaluate_coverage_package_writers.py \
  --package-dir llm_memory_validation/natural_adjudicated_100_gemini_flash/coverage_package \
  --out-dir llm_memory_validation/natural_adjudicated_100_gemini_flash/writer_adapters \
  --budgets 30,60,100 \
  --methods opt,oracle_gvt,memgpt_tiered,amem_graph,mem0_extract,amac_admission,estimated_gvt,density_only,summary_only,fact_only,recency_raw

Expected outputs:

• llm_memory_validation/natural_adjudicated_100_gemini_flash/writer_adapters/raw_results.jsonl
• llm_memory_validation/natural_adjudicated_100_gemini_flash/writer_adapters/summary.json
• llm_memory_validation/natural_adjudicated_100_gemini_flash/writer_adapters/summary.md
• llm_memory_validation/natural_adjudicated_100_gemini_flash/writer_adapters/REPORT.md
• llm_memory_validation/natural_adjudicated_100_gemini_flash/writer_adapters/run_manifest.json

Current run: 87 accepted/corrected adjudicated examples, zero API calls. Letta/MemGPT-style reaches 0.638/0.433/0.431, A-Mem-style reaches 0.481/0.374/0.377, and density-only reaches 0.991/0.955/0.962 at budgets 30/60/100. The density result is a warning that this copied-candidate natural denominator is unusually density-friendly.

Human-Edited Writer Adapter Diagnostic

This local run scores the same Letta/MemGPT-style, A-Mem-style, Mem0-style, and A-MAC-style adapters on the exported human-edited coverage package. It is a zero-API denominator-matched check. It does not run the Letta service or MemGPT controller; the checked-out Letta repository requires a service/API/model configuration for a true production run.

python llm_memory_validation/evaluate_coverage_package_writers.py \
  --package-dir llm_memory_validation/human_style_examples/coverage_package \
  --out-dir llm_memory_validation/human_style_examples/writer_adapters \
  --budgets 150,300,600,1000 \
  --methods opt,oracle_gvt,memgpt_tiered,amem_graph,mem0_extract,amac_admission,estimated_gvt,density_only,summary_only,fact_only,recency_raw

Expected outputs:

• llm_memory_validation/human_style_examples/writer_adapters/raw_results.jsonl
• llm_memory_validation/human_style_examples/writer_adapters/summary.json
• llm_memory_validation/human_style_examples/writer_adapters/summary.md
• llm_memory_validation/human_style_examples/writer_adapters/REPORT.md
• llm_memory_validation/human_style_examples/writer_adapters/run_manifest.json

Current run: 85 query-resolved examples, zero API calls. Letta/MemGPT-style reaches 0.847, A-Mem-style reaches 0.876, Mem0-style reaches 0.753, and A-MAC-style reaches 0.835 across budgets 150/300/600/1000. Density-only is 1.000 on this per-query exported package, so this row is a MemGPT-style adapter reproducibility check rather than the strongest algorithmic separation.

Faithful MemGPT/Letta Union Baseline

This no-API runner is the current MemGPT/Letta-strengthened baseline on the adjudicated natural package. It checks the local external_repos/letta checkout metadata, records that the actual Letta import path is not available without the full service dependency stack, then simulates the relevant core/archival/recall memory tiers over exported package candidates. Writing and retrieval use visible metadata only; hidden coverage is used only for scoring, except in the analysis-only upper-pruned bound row.

python llm_memory_validation/run_faithful_memgpt_letta_baseline.py \
  --package-dir llm_memory_validation/natural_adjudicated_100_gemini_flash/coverage_package \
  --out-dir llm_memory_validation/natural_adjudicated_100_gemini_flash/faithful_memgpt_letta_union \
  --budgets 30,60,100 \
  --limit 87

Expected outputs:

• llm_memory_validation/natural_adjudicated_100_gemini_flash/faithful_memgpt_letta_union/raw_results.jsonl
• llm_memory_validation/natural_adjudicated_100_gemini_flash/faithful_memgpt_letta_union/summary.json
• llm_memory_validation/natural_adjudicated_100_gemini_flash/faithful_memgpt_letta_union/REPORT.md
• llm_memory_validation/natural_adjudicated_100_gemini_flash/faithful_memgpt_letta_union/written_stores.jsonl
• llm_memory_validation/natural_adjudicated_100_gemini_flash/faithful_memgpt_letta_union/run_manifest.json

Current run: 87/87 examples, zero API calls. Archival-search pruning reaches 0.746/0.739/0.866 ratio to union OPT at budgets 30/60/100; recency pruning reaches 0.642/0.700/0.877; the analysis-only upper-pruned bound reaches 0.829/0.907/0.939.

Actual Letta OpenRouter Passage Run

This runs the checked-out Letta server (external_repos/letta, version 0.16.7) with Postgres/pgvector, OpenRouter Gemini, and authenticated OpenRouter passage embeddings. Apply llm_memory_validation/patches/letta_openrouter_embedding_auth.patch to the Letta checkout before starting the server; without it, OpenRouter passage search uses the wrong API key path.

.\.venv_letta_prod\Scripts\python.exe llm_memory_validation\run_actual_letta_openrouter_baseline.py `
  --package-dir llm_memory_validation\natural_adjudicated_100_gemini_flash\coverage_package `
  --out-dir llm_memory_validation\natural_adjudicated_100_gemini_flash\actual_letta_openrouter_gemini_passage_87 `
  --limit 87 `
  --budgets 30,60,100 `
  --include-salience-pruned `
  --include-oracle-pruned-upper `
  --max-steps 12 `
  --message-retries 2 `
  --request-sleep 0.02

Expected outputs:

• llm_memory_validation/natural_adjudicated_100_gemini_flash/actual_letta_openrouter_gemini_passage_87/raw_results.jsonl
• llm_memory_validation/natural_adjudicated_100_gemini_flash/actual_letta_openrouter_gemini_passage_87/summary.json
• llm_memory_validation/natural_adjudicated_100_gemini_flash/actual_letta_openrouter_gemini_passage_87/REPORT.md
• llm_memory_validation/natural_adjudicated_100_gemini_flash/actual_letta_openrouter_gemini_passage_87/written_stores.jsonl
• llm_memory_validation/natural_adjudicated_100_gemini_flash/actual_letta_openrouter_gemini_passage_87/coverage_scoring_calls.jsonl
• llm_memory_validation/natural_adjudicated_100_gemini_flash/actual_letta_openrouter_gemini_passage_87/salience_scoring_calls.jsonl

Current run: 87/87 examples, zero failed instances. Letta writes archival passages for 85 examples and core-memory atoms for 30 examples. The combined core+archival store reaches union-OPT ratios 0.652/0.696/0.734 with salience pruning, 0.219/0.260/0.342 with recency pruning, and 0.723/0.763/0.765 for the analysis-only upper-pruned bound at budgets 30/60/100.

Actual A-Mem Gemini-Flash Pilot

This runs the checked-out public external_repos/AgenticMemory implementation, using Gemini Flash through OpenRouter for A-Mem metadata/evolution calls and for post-hoc coverage scoring. It reports a finite union denominator over package candidates plus A-Mem-written memories. The full-memory rows score A-Mem's actual stored notes; the metadata rows are a compact diagnostic serialization of A-Mem-generated context/keywords/tags/links.

python llm_memory_validation/run_actual_amem_natural_baseline.py \
  --package-dir llm_memory_validation/natural_adjudicated_100_gemini_flash/coverage_package \
  --out-dir llm_memory_validation/natural_adjudicated_100_gemini_flash/actual_amem_gemini_flash_87 \
  --limit 87 \
  --budgets 30,60,100,5000 \
  --amem-model google/gemini-2.5-flash \
  --coverage-model google/gemini-2.5-flash \
  --request-sleep 0.02 \
  --amem-max-tokens 3000

Expected outputs:

• llm_memory_validation/natural_adjudicated_100_gemini_flash/actual_amem_gemini_flash_87/REPORT.md
• llm_memory_validation/natural_adjudicated_100_gemini_flash/actual_amem_gemini_flash_87/summary.json
• llm_memory_validation/natural_adjudicated_100_gemini_flash/actual_amem_gemini_flash_87/raw_results.jsonl
• llm_memory_validation/natural_adjudicated_100_gemini_flash/actual_amem_gemini_flash_87/written_stores.jsonl
• llm_memory_validation/natural_adjudicated_100_gemini_flash/actual_amem_gemini_flash_87/coverage_scoring_calls.jsonl
• llm_memory_validation/natural_adjudicated_100_gemini_flash/actual_amem_gemini_flash_87/run_manifest.json

Current 87-example run: raw full A-Mem notes have mean serialized cost 4446 words and therefore score 0.000/0.000/0.000 at budgets 30/60/100; at the diagnostic budget 5000, the raw full-store oracle upper reaches 0.845. The compact metadata diagnostic has mean cost 66 words and reaches 0.204/0.158/0.180 with oracle pruning at budgets 30/60/100. The run used 524 cached API prompts, 2,433,021 tokens, and an estimated OpenRouter cost of $1.576.

Actual Mem0 Smoke

This verifies executable integration with the public Mem0 codebase. It is not a benchmark and should not be reported as a budget-matched Mem0 comparison.

Prerequisites from this environment:

python -m pip install qdrant-client==1.12.2 rank-bm25==0.2.2 litellm==1.83.7
python -m pip install -e external_repos/mem0
python -m pip install "huggingface-hub>=0.34,<1.0"

Run:

python llm_memory_validation/mem0_actual_smoke.py \
  --api-env api.env \
  --out-dir llm_memory_validation/mem0_actual_smoke

Expected outputs:

• llm_memory_validation/mem0_actual_smoke/search_result.json
• llm_memory_validation/actual_system_repo_audit/REPORT.md

LongMemEval-S Retrieval Transfer

Diagnostic only after the 9-page compression pass. This report is retrieval-only: no answer generation, no abstention scoring, and no exact OPT denominator.

To regenerate the focus report from the cached retrieval rows:

python llm_memory_validation/longmemeval_focus_report.py \
  --summary-json llm_memory_validation/competitor_run_v2/summary.json \
  --retrieval-rows-json llm_memory_validation/competitor_run_v2/retrieval_rows.json \
  --output-dir llm_memory_validation/longmemeval_focus_report_core4 \
  --methods dense_budgeted_bsc,dense_rag_e5,dense_budgeted_replay,fifo_replay

Expected outputs:

• llm_memory_validation/longmemeval_focus_report_core4/summary.json
• llm_memory_validation/longmemeval_focus_report_core4/REPORT.md

The current paper-facing label map is:

• dense_budgeted_bsc: MemAudit writer + dense retrieval
• dense_rag_e5: Full raw-store dense retrieval
• dense_budgeted_replay: Budgeted raw replay + dense retrieval
• fifo_replay: FIFO raw replay

To regenerate the upstream dense retrieval rows, use:

python llm_memory_validation/paper_competitor_suite.py \
  --output-dir llm_memory_validation/competitor_run_v2 \
  --topk 5 \
  --retriever-model intfloat/e5-base-v2

This upstream regeneration downloads external data/model artifacts and may vary with model or dataset revisions unless those are pinned outside this repository.

GPT-5.5 Frozen-Context Reader

Appendix diagnostic only after the 9-page compression pass. The current artifact uses frozen top-5 retrieval contexts, openai/gpt-5.5 through OpenRouter, and the answer_if_supported prompt.

Set up api.env locally. Do not commit it.

OPENROUTER_API_KEY=...

Then run:

python llm_memory_validation/longmemeval_reader_eval.py \
  --dataset-json llm_memory_validation/cache/longmemeval_s_cleaned.json \
  --retrieval-rows-json llm_memory_validation/competitor_run_v2/retrieval_rows.json \
  --output-dir llm_memory_validation/longmemeval_reader_api_gpt55_answer_supported_focus_full \
  --methods dense_budgeted_bsc,dense_rag_e5,dense_budgeted_replay,fifo_replay \
  --focus-only \
  --focus-types knowledge-update,temporal-reasoning \
  --reader openrouter \
  --reader-model openai/gpt-5.5 \
  --prompt-style answer_if_supported \
  --api-env api.env \
  --api-cache llm_memory_validation/openrouter_cache_gpt55_answer_supported_focus_full.json

Expected outputs:

• llm_memory_validation/longmemeval_reader_api_gpt55_answer_supported_focus_full/summary.json
• llm_memory_validation/longmemeval_reader_api_gpt55_answer_supported_focus_full/REPORT.md
• llm_memory_validation/longmemeval_reader_api_gpt55_answer_supported_focus_full/reader_outputs.jsonl
• llm_memory_validation/longmemeval_reader_api_gpt55_answer_supported_focus_full/predictions.json

The committed/cacheable outputs should be treated as the reproducible artifact for the paper. Re-running the API may change costs, latency, or model behavior.

Reader Audit

Appendix diagnostic only after the 9-page compression pass.

python llm_memory_validation/longmemeval_reader_eval.py \
  --analyze-errors \
  --run-dir llm_memory_validation/longmemeval_reader_api_gpt55_answer_supported_focus_full

Expected outputs in the same run directory:

• ERROR_AUDIT.md
• error_audit_summary.json
• error_audit_rows.jsonl
• failure_examples.jsonl
• semantic_audit_sample_50.jsonl
• normalized_scoring.json
• llm_memory_validation/scoring_audit_gpt55/normalized_scoring_v2.json

Deterministic Decomposition

Diagnostic only after the 9-page compression pass. This is a local evidence-only reader path and does not use an API.

python run_oraclemem_mvp.py \
  --n-seeds 300 \
  --budgets 0.05,0.10,0.20 \
  --distribution base,update_chain,temporal_interval \
  --methods opt,oracle_gvt,density_only,greedy,recency_raw,reservoir_raw,summary_only,fact_only,no_tombstone_gvt \
  --enable-retrieval \
  --retrieval fixed,oracle \
  --reader evidence_only \
  --out oraclemem_runs/decomp_det_300

Expected outputs:

• oraclemem_runs/decomp_det_300/raw_results.jsonl
• oraclemem_runs/decomp_det_300/summary.json
• oraclemem_runs/decomp_det_300/summary.md

MILP Verification

Referenced in the exact-small solver audit text. This optional run requires pulp from requirements-milp.txt.

python run_oraclemem_mvp.py \
  --n-seeds 100 \
  --budgets 0.02,0.05,0.10,0.20 \
  --distribution base,update_chain,temporal_interval \
  --methods opt \
  --solver milp \
  --verify-against exact_stdlib \
  --out oraclemem_runs/milp_verify_100_agent4

Expected outputs:

• oraclemem_runs/milp_verify_100_agent4/raw_results.jsonl
• oraclemem_runs/milp_verify_100_agent4/summary.json
• oraclemem_runs/milp_verify_100_agent4/summary.md
• oraclemem_runs/milp_verify_100_agent4/REPORT.md

Gemini Flash-Lite Diagnostic

This API run is a robustness diagnostic, not a theorem-facing result. It uses OpenRouter model google/gemini-3.1-flash-lite-preview and requires api.env.

python llm_memory_validation/longmemeval_reader_eval.py \
  --reader openrouter \
  --reader-model google/gemini-3.1-flash-lite-preview \
  --prompt-style answer_if_supported \
  --focus-only \
  --methods dense_budgeted_bsc,fifo_replay \
  --api-env api.env \
  --api-cache llm_memory_validation/openrouter_cache_gemini31_flash_lite_focus_full_bsc_fifo.json \
  --output-dir llm_memory_validation/longmemeval_reader_api_gemini31_flash_lite_focus_full_bsc_fifo \
  --api-max-tokens 320 \
  --api-timeout 120 \
  --temperature 0 \
  --request-sleep 0.02 \
  --bootstrap 1000 \
  --save-prompts

Noisy Estimated-Policy Diagnostic

This run does not call an API. It records Gemini Flash-Lite as provenance for a local noisy estimated-utility profile and is useful as a synthetic stress diagnostic for non-oracle writer evaluation.

python run_oraclemem_mvp.py \
  --n-seeds 500 \
  --distribution scope_shift_v2,density_trap_v2 \
  --budgets 4,6 \
  --methods opt,oracle_gvt,estimated_gvt,estimated_utility,mem0_extract,amac_admission,no_tombstone_gvt,no_tombstone_opt \
  --estimated-model google/gemini-3.1-flash-lite-preview \
  --estimated-profile noisy_gemini_flash_lite_v1 \
  --enable-retrieval \
  --retrieval fixed,oracle \
  --export-coverage-matrices \
  --coverage-package-limit 4 \
  --out-dir oraclemem_runs/estimated_policy_noisy_noapi_1000

To audit an exported coverage package:

python scripts/audit_coverage_artifacts.py \
  --no-defaults \
  --artifact exported_oraclemem_package=oraclemem_runs/estimated_policy_noisy_noapi_1000/coverage_instances/scope_shift_v2/seed_0 \
  --output-dir oraclemem_runs/estimated_policy_noisy_noapi_1000/coverage_audit

Train/Dev Estimated-Writer Diagnostic

This local run trains a ridge utility estimator on synthetic train seeds and evaluates estimated_* methods only on held-out dev seeds. It does not call an API and is diagnostic rather than final deployed-writer evidence.

python run_oraclemem_mvp.py \
  --n-seeds 60 \
  --train-dev-estimator \
  --train-fraction 0.5 \
  --distribution base,update_chain,temporal_interval,density_trap,scope_shift,summary_tradeoff,redundancy_heavy,abstention_hard,scope_shift_v2,density_trap_v2 \
  --budgets 4,6 \
  --methods opt,oracle_gvt,estimated_gvt,estimated_utility,mem0_extract,amac_admission,no_tombstone_gvt,no_tombstone_opt \
  --out-dir oraclemem_runs/estimated_policy_train_dev_local_60

Known Non-Reproducible Or External Pieces

• Local LaTeX compilation depends on a TeX distribution; this machine did not have latexmk, pdflatex, or tectonic on PATH.
• GPT-5.5 reader outputs require OpenRouter access, model availability, and API spending. Use the cached reader outputs for paper auditability.
• Gemini natural coverage and actual Mem0 smoke outputs require OpenRouter access if regenerated from scratch; use cached artifacts for audit where possible.
• LongMemEval-S retrieval regeneration downloads the dataset and intfloat/e5-base-v2; exact rows can drift if upstream artifacts change.
• API costs in summary.json are historical and should not be treated as a stable price quote.