README.md

metadata

license: mit
pretty_name: MemAudit Code Artifact
tags:
  - llm-memory
  - benchmark
  - evaluation
  - reproducibility

MemAudit

MemAudit is an exact-oracle evaluation protocol for budgeted long-term LLM memory writing. The core question is finite and package-conditional:

Given a fixed storage budget and a finite semantic evidence package, how close is a written memory store to the best package-feasible store?

This repository contains the manuscript, exact-small synthetic benchmarks, validity-heavy stress benchmarks, natural support-sliced coverage packages, Mem0 diagnostic rescoring artifacts, and reproducibility scripts.

MemAudit is not a runtime memory product. It is an evaluation layer for memory writers: it scores finite candidate packages, budgeted representation choices, and external written stores against explicit denominators.

Quickcheck

Run the deterministic tests:

python -m unittest test_oraclemem.py

Run a tiny exact-oracle smoke benchmark:

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

Expected smoke outputs:

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

Main Artifacts

• main.tex: active manuscript.
• references.bib: bibliography.
• figures/: paper figure assets generated from cached experiment summaries.
• oraclemem_runs/exact_500: exact-small 500-instance sweep.
• oraclemem_runs/stress_exact_500: validity-heavy stress sweep.
• oraclemem_runs/representative_writers_500: non-oracle writer diagnostic sweep with Estimated-GVT and A-MAC-like admission.
• llm_memory_validation/oraclemem_natural_200_gemini_v2: Natural-200 support-sliced coverage package.
• llm_memory_validation/natural_adjudicated_100_gemini_flash: stricter adjudicated natural subset.
• llm_memory_validation/natural_spotcheck_30_gemini31_flash_lite: independent Gemini Flash-Lite adjudication spot-check.
• llm_memory_validation/human_style_examples: 100 fictional human-edited/audited natural examples, exported coverage package, exact package evaluation, and actual A-Mem run.
• llm_memory_validation/human_style_examples/learned_writer_transfer: coverage-blind learned writer transfer diagnostic trained on synthetic plus Natural-200 labels and tested on the human-edited package.
• llm_memory_validation/human_style_examples/learned_writer_transfer_synth_only and llm_memory_validation/human_style_examples/learned_writer_transfer_natural_only: training-source ablations for the learned writer transfer diagnostic.
• llm_memory_validation/human_style_examples/writer_adapters: denominator-matched Letta/MemGPT-style, A-Mem-style, Mem0-style, and A-MAC-style adapter diagnostics on the exported human-edited coverage package.
• llm_memory_validation/natural_adjudicated_100_gemini_flash/writer_adapters: denominator-matched Letta/MemGPT-style and A-Mem-style adapter diagnostics on the adjudicated natural package.
• llm_memory_validation/natural_adjudicated_100_gemini_flash/faithful_memgpt_letta_union: no-API faithful MemGPT/Letta core/archival baseline scored with a package-plus-written-store union denominator.
• llm_memory_validation/natural_adjudicated_100_gemini_flash/actual_letta_openrouter_gemini_passage_87: executable Letta server run on 87 adjudicated examples with OpenRouter Gemini, authenticated OpenRouter passage embeddings, archival-memory tools, and the union denominator.
• llm_memory_validation/mem0_rescore_adjudicated100_gemini_flash: Mem0 diagnostic rescoring on the adjudicated subset.
• llm_memory_validation/natural_adjudicated_100_gemini_flash/actual_amem_gemini_flash_87: executable public A-Mem run on 87 adjudicated examples using Gemini Flash and the union denominator.
• llm_memory_validation/human_style_examples/actual_amem_gemini_flash_100: executable public A-Mem run on the human-edited package using Gemini Flash and the union denominator.

See artifact_manifest.md for table-to-artifact mapping and full rerun commands. See REPRODUCIBILITY.md for setup, exact-oracle runs, API runs, and known local build limitations.

Denominator Types

• Package ratio: exact ratio to OPT_P(B) for a finite MemAudit candidate package.
• Union ratio: exact ratio to OPT_{P^+(Y)}(B) after adding an external written store to the candidate package.
• Upper-pruned bound: best budget-feasible subset of an external store, used only to separate extraction quality from budget-aware selection.
• Retrieval/reader metrics: downstream diagnostics, not MemAudit optimum ratios.

Caveats

The strongest exact claims are finite-package claims. LongMemEval-derived natural coverage packages are model-adjudicated; the separate human_style_examples package is human-edited/audited but does not include an inter-annotator agreement file. LongMemEval reader/retrieval results are downstream diagnostics and do not have exact OPT denominators. Mem0 and A-Mem rescoring use union-denominator and upper-pruned-bound diagnostics rather than claiming deployable optimal pruning policies.

Do not commit API keys. api.env is local-only and should stay ignored.