| Exact-small 500 |
fig:exact-budget-sweep |
oraclemem_runs/exact_500 |
summary.md, summary.json, raw_results.jsonl, optional coverage_instances/base/seed_*/coverage_matrix.jsonl |
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 --export-coverage-matrices |
| Validity-heavy stress 500 |
fig:stress-validity |
oraclemem_runs/stress_exact_500 |
summary.md, summary.json, raw_results.jsonl, optional coverage_instances/<distribution>/seed_*/coverage_matrix.jsonl |
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 --export-coverage-matrices |
| Representative non-oracle writers |
Main text diagnostic |
oraclemem_runs/representative_writers_500 |
summary.md, summary.json, raw_results.jsonl |
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 |
| No-API proxy writer diagnostic |
Diagnostic only |
oraclemem_runs/proxy_writer_baselines_50 |
REPORT.md, summary.md, summary.json, raw_results.jsonl |
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 |
| Gemini Natural-200 coverage package |
tab:natural-reliability |
llm_memory_validation/oraclemem_natural_200_gemini_v2 |
REPORT.md, summary.json, coverage_resolved_summary.json, coverage_resolution_report.md, coverage_package/, coverage_audit/REPORT.md |
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; then 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 |
| Actual Mem0 Natural-200 baseline |
tab:natural-adjudicated |
llm_memory_validation/mem0_natural200_actual |
REPORT.md, summary.json, raw_results.jsonl, written_stores.jsonl, coverage_scoring_calls.jsonl; cloned repo external_repos/mem0 |
python llm_memory_validation/run_mem0_natural_baseline.py --package-dir llm_memory_validation/oraclemem_natural_200_gemini_v2/coverage_package --out-dir llm_memory_validation/mem0_natural200_actual --limit 200 --budgets 30,60,100 --include-oracle-pruned-upper |
| Secondary natural annotation audit |
Limitation/adjudication warning |
llm_memory_validation/natural50_annotation_agreement_gemini31_vs_gemini25 |
REPORT.md, summary.json, agreement_rows.jsonl; secondary package llm_memory_validation/natural50_secondary_gemini25 |
python llm_memory_validation/gemini_natural_oraclemem.py --model google/gemini-2.5-flash-lite --limit 50 --distractors-per-example 0 --max-session-words 1800 --budgets 30,60,100 --out-dir llm_memory_validation/natural50_secondary_gemini25 --request-sleep 0.02; then python llm_memory_validation/compare_natural_coverage_annotations.py --primary llm_memory_validation/oraclemem_natural_200_gemini_v2/coverage_package --secondary llm_memory_validation/natural50_secondary_gemini25/coverage_package --out-dir llm_memory_validation/natural50_annotation_agreement_gemini31_vs_gemini25 |
| Gemini Flash adjudicated natural subset |
tab:natural-adjudicated package rows |
llm_memory_validation/natural_adjudicated_100_gemini_flash |
REPORT.md, adjudication_summary.json, summary.json, coverage_package/, coverage_audit/REPORT.md |
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; then python scripts/audit_coverage_artifacts.py --no-defaults --artifact natural_adjudicated_100_gemini_flash=llm_memory_validation/natural_adjudicated_100_gemini_flash/coverage_package --output-dir llm_memory_validation/natural_adjudicated_100_gemini_flash/coverage_audit |
| Gemini Flash-Lite spot-check |
tab:natural-reliability spot-check row |
llm_memory_validation/natural_spotcheck_30_gemini31_flash_lite |
REPORT.md, adjudication_summary.json, summary.json, coverage_package/, coverage_audit/REPORT.md |
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; then 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 |
| Human-edited natural seed package |
tab:natural-reliability human row |
llm_memory_validation/human_style_examples |
examples_100.jsonl, README.md, eval_package_100/REPORT.md, eval_package_100/summary.json, eval_package_100/raw_results.jsonl |
python scripts/validate_human_style_examples.py llm_memory_validation/human_style_examples/examples_100.jsonl; then 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 |
| Human-edited coverage-package export |
Human-edited/A-Mem paragraph |
llm_memory_validation/human_style_examples/coverage_package and llm_memory_validation/human_style_examples/coverage_package_audit |
coverage_package/*.jsonl, coverage_package/candidate_generation_manifest.json, coverage_package_audit/REPORT.md, coverage_package_audit/summary.json |
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; then 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 |
| Human-edited writer adapters |
System-style adapter paragraph |
llm_memory_validation/human_style_examples/writer_adapters |
REPORT.md, summary.md, summary.json, raw_results.jsonl, run_manifest.json |
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 |
| Actual A-Mem on human-edited package |
Human-edited/A-Mem paragraph |
llm_memory_validation/human_style_examples/actual_amem_gemini_flash_100 |
REPORT.md, summary.json, raw_results.jsonl, written_stores.jsonl, coverage_scoring_calls.jsonl, run_manifest.json; cloned repo external_repos/AgenticMemory |
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 |
| Learned writer transfer diagnostic |
Learned-writer paragraph |
llm_memory_validation/human_style_examples/learned_writer_transfer |
REPORT.md, summary.md, summary.json, raw_results.jsonl, train_manifest.json |
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 |
| Learned writer source ablations |
Learned-writer paragraph |
llm_memory_validation/human_style_examples/learned_writer_transfer_synth_only, llm_memory_validation/human_style_examples/learned_writer_transfer_natural_only |
REPORT.md, summary.md, summary.json, raw_results.jsonl, train_manifest.json |
Synthetic-only: 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; Natural-only: 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 |
| Natural writer adapters |
System-style adapter paragraph |
llm_memory_validation/natural_adjudicated_100_gemini_flash/writer_adapters |
REPORT.md, summary.md, summary.json, raw_results.jsonl, run_manifest.json |
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 |
| Faithful MemGPT/Letta union baseline |
tab:natural-adjudicated MemGPT/Letta rows and paragraph |
llm_memory_validation/natural_adjudicated_100_gemini_flash/faithful_memgpt_letta_union |
REPORT.md, summary.json, raw_results.jsonl, written_stores.jsonl, run_manifest.json; cloned repo external_repos/letta |
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 |
| Actual Letta OpenRouter passage run |
Actual Letta paragraph |
llm_memory_validation/natural_adjudicated_100_gemini_flash/actual_letta_openrouter_gemini_passage_87 |
REPORT.md, summary.json, raw_results.jsonl, written_stores.jsonl, coverage_scoring_calls.jsonl, salience_scoring_calls.jsonl; cloned repo external_repos/letta; patch llm_memory_validation/patches/letta_openrouter_embedding_auth.patch |
With Letta server running on Postgres/pgvector and the OpenRouter embedding-auth patch applied: .\.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 |
| Mem0 rescore on adjudicated subset |
tab:natural-adjudicated Mem0 rows |
llm_memory_validation/mem0_rescore_adjudicated100_gemini_flash |
REPORT.md, summary.json, raw_results.jsonl, coverage_scoring_calls.jsonl, salience_scoring_calls.jsonl |
python llm_memory_validation/score_mem0_written_stores.py --package-dir llm_memory_validation/natural_adjudicated_100_gemini_flash/coverage_package --written-stores-jsonl llm_memory_validation/mem0_natural200_actual/written_stores.jsonl --out-dir llm_memory_validation/mem0_rescore_adjudicated100_gemini_flash --coverage-model google/gemini-2.5-flash --salience-model google/gemini-2.5-flash --budgets 30,60,100 --include-salience-pruned --include-oracle-pruned-upper --request-sleep 0.02 |
| Actual A-Mem Gemini-Flash run |
Actual A-Mem paragraph and tab:natural-adjudicated A-Mem rows |
llm_memory_validation/natural_adjudicated_100_gemini_flash/actual_amem_gemini_flash_87 |
REPORT.md, summary.json, raw_results.jsonl, written_stores.jsonl, coverage_scoring_calls.jsonl, run_manifest.json; cloned repo external_repos/AgenticMemory |
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 |
| Actual Mem0 smoke |
Superseded external validation smoke |
llm_memory_validation/mem0_actual_smoke and llm_memory_validation/actual_system_repo_audit |
search_result.json, actual_system_repo_audit/REPORT.md; cloned repo external_repos/mem0 |
python llm_memory_validation/mem0_actual_smoke.py --api-env api.env --out-dir llm_memory_validation/mem0_actual_smoke |
| Train/dev estimated-writer diagnostic |
Diagnostic only |
oraclemem_runs/estimated_policy_train_dev_local_60 |
summary.md, summary.json, raw_results.jsonl |
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 |
| LongMemEval-S retrieval transfer |
Appendix diagnostic |
llm_memory_validation/longmemeval_focus_report_core4 |
REPORT.md, summary.json; upstream llm_memory_validation/competitor_run_v2/retrieval_rows.json |
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 |
| GPT-5.5 frozen-context reader |
Appendix diagnostic |
llm_memory_validation/longmemeval_reader_api_gpt55_answer_supported_focus_full |
REPORT.md, summary.json, reader_outputs.jsonl, predictions.json |
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 |
| Conditional reader audit |
Appendix diagnostic |
llm_memory_validation/longmemeval_reader_api_gpt55_answer_supported_focus_full |
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 |
python llm_memory_validation/longmemeval_reader_eval.py --analyze-errors --run-dir llm_memory_validation/longmemeval_reader_api_gpt55_answer_supported_focus_full |
| LongMemEval-S cached diagnostic check |
Claim-boundary/reporting aid |
llm_memory_validation/longmemeval_cached_diagnostic_check |
REPORT.md, summary.json; reads the cached retrieval, GPT-5.5, Gemini, GPT-5.4-mini, scoring-audit, failure-audit, and prompt-dev summaries |
python llm_memory_validation/longmemeval_cached_diagnostic_check.py |
| Deterministic decomposition |
Appendix diagnostic |
oraclemem_runs/decomp_det_300 |
summary.md, summary.json, raw_results.jsonl |
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 |
| MILP solver audit text |
Exact-small solver paragraph |
oraclemem_runs/milp_verify_100_agent4 |
REPORT.md, summary.md, summary.json, raw_results.jsonl |
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 |
| Gemini 3.1 Flash-Lite reader diagnostic |
Appendix/API robustness only |
llm_memory_validation/longmemeval_reader_api_gemini31_flash_lite_focus_full_bsc_fifo |
summary.json, REPORT.md, reader_outputs.jsonl, predictions.json |
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 |
Diagnostic only |
oraclemem_runs/estimated_policy_noisy_noapi_1000 |
summary.md, summary.json, raw_results.jsonl, coverage_instances/, coverage_audit_final/summary.json |
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 |
| Deterministic estimated-policy diagnostic |
Superseded diagnostic |
oraclemem_runs/estimated_policy_gemini31_flash_lite_1000 |
summary.md, summary.json, raw_results.jsonl |
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 gemini_flash_lite_v1 --enable-retrieval --retrieval fixed,oracle --out-dir oraclemem_runs/estimated_policy_gemini31_flash_lite_1000 |
| Paper figures |
fig:* |
figures/ |
Canonical summaries listed above |
python scripts/make_figures.py |
