| # MemAudit Evaluation Card |
|
|
| ## Intended Use |
|
|
| MemAudit evaluates long-term LLM memory writers under an explicit storage |
| budget and a finite set of candidate memories. It is intended for measuring |
| write-time memory quality, comparing budgeted representation choices, auditing |
| validity-state/tombstone behavior, and diagnosing whether external memory stores |
| fail through extraction quality or budget-aware selection. |
|
|
| ## Not Intended Use |
|
|
| MemAudit ratios are not end-to-end assistant quality guarantees. They are not |
| global optima over all possible memories, all possible natural-language |
| compressions, or all possible retrieval policies. LongMemEval reader/retrieval |
| numbers are downstream diagnostics, not exact oracle ratios. |
|
|
| ## Denominators |
|
|
| - Package denominator: `OPT_P(B)`, the exact optimum for a finite MemAudit |
| package. |
| - Union denominator: `OPT_{P^+(Y)}(B)`, the exact optimum after adding an |
| external written store `Y` to the package candidate set. |
| - Upper-pruned bound: the best budget-feasible subset of an external store, |
| used only as an extraction-versus-selection diagnostic. |
| - Retrieval/reader metrics: accuracy, recall, F1, abstention, stale-answer rate, |
| and token cost; no exact OPT denominator is claimed. |
|
|
| ## Main Package Artifacts |
|
|
| - Synthetic exact-small: `oraclemem_runs/exact_500`. |
| - Validity-heavy stress: `oraclemem_runs/stress_exact_500`. |
| - Representative non-oracle writers: `oraclemem_runs/representative_writers_500`. |
| - Natural support-sliced package: `llm_memory_validation/oraclemem_natural_200_gemini_v2`. |
| - Natural adjudicated subset: `llm_memory_validation/natural_adjudicated_100_gemini_flash`. |
| - Natural Flash-Lite spot-check: `llm_memory_validation/natural_spotcheck_30_gemini31_flash_lite`. |
| - Human-edited natural seed package: `llm_memory_validation/human_style_examples`. |
| - Learned writer transfer diagnostic: `llm_memory_validation/human_style_examples/learned_writer_transfer`. |
| - Natural writer adapters: `llm_memory_validation/natural_adjudicated_100_gemini_flash/writer_adapters`. |
| - Mem0 adjudicated rescore: `llm_memory_validation/mem0_rescore_adjudicated100_gemini_flash`. |
|
|
| ## Annotation Status |
|
|
| Exact synthetic coverage matrices are generated from the simulator and are |
| machine-checkable. Natural coverage packages are model-generated and |
| model-adjudicated; they are useful reliability diagnostics but have not |
| undergone human audit. The secondary natural audit showed that unsupported |
| natural annotations are a bottleneck, while the 30-example Gemini Flash-Lite |
| spot-check provides an additional model-adjudicated consistency check. The |
| `human_style_examples` package has been human-edited/audited and is structurally |
| validated, but it does not include independent inter-annotator agreement. The |
| paper therefore treats Natural-200 and the human-edited package as reliability |
| and artifact-validity evidence rather than definitive natural ground truth. |
|
|
| ## External Memory Systems |
|
|
| External stores such as Mem0 are evaluated with union-denominator diagnostics. |
| This prevents the invalid claim that an external writer should be measured |
| against a denominator that excludes its own candidate memories. The upper-pruned |
| upper is not a deployable method; it asks how much value is present in the |
| written store if budget selection were solved post hoc. |
|
|
| System-style local adapters such as Letta/MemGPT-style tiering and A-Mem-style |
| graph writing are evaluated as visible-metadata policies over package |
| candidates. They are denominator-matched baselines, not full published-system |
| executions. The checked-out Letta repository was inspected, but a true |
| Letta/MemGPT run requires a service/API/model configuration; the reported |
| MemGPT-style rows therefore remain local adapter rows. |
|
|
| The actual A-Mem run executes the checked-out public `AgenticMemory` code path |
| on the 87-example adjudicated package with Gemini Flash. It is intentionally labeled separately from the local |
| adapter rows: raw A-Mem notes are scored as full external memories, and a compact |
| metadata view is reported only as a diagnostic derived from A-Mem's generated |
| context/keywords/tags/links. |
|
|
| The human-edited examples are also exported to the same coverage-package schema |
| and used for an actual A-Mem run. That result is stronger than a purely |
| model-adjudicated package, but it remains a sanity check rather than an |
| inter-annotator benchmark because the examples are fictional and short. |
| The exported human package also has zero-API system-adapter rows: the |
| Letta/MemGPT-style adapter reaches 0.847 ratio to exact package OPT, while |
| density-only is 1.000, so this row is treated as a protocol check rather than a |
| separation result. |
|
|
| The adjudicated natural package includes a stronger faithful MemGPT/Letta union |
| baseline. It simulates core/archival/recall memory tiers over package-derived |
| written memories and scores against a package-plus-written-store union |
| denominator. It is still not a Letta server/API run, but it is closer to the |
| MemGPT memory architecture than the simple adapter. |
|
|
| ## API Use |
|
|
| API calls are used for natural package construction, adjudication, external |
| store rescoring, and reader diagnostics. Exact synthetic labels and exact |
| synthetic optima are deterministic and do not depend on API calls. API costs and |
| cache files are recorded in the corresponding run directories. |
|
|
| ## Learned Writer Status |
|
|
| The learned writer transfer diagnostic trains a local visible-feature estimator |
| from oracle labels on train packages, then evaluates held-out selections without |
| access to hidden coverage labels or query requirements. It is a deployable-writer |
| diagnostic, not a proof that learned writing is solved across natural traces. |
| The source ablations show that the current paper-facing estimator depends on |
| combining synthetic stress labels with Natural-200 labels; neither source alone |
| is sufficient on the human-edited package. |
|
|
| ## Quickcheck |
|
|
| ```powershell |
| 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 |
| ``` |
|
|
| ## Release Checks |
|
|
| - Verify no API keys or private credentials are included. |
| - Verify paper-facing labels match `artifact_manifest.md`. |
| - Verify no natural package is described as human-validated unless a human audit |
| has actually been run. |
| - Verify greedy, retrieval, and reader diagnostics are not labeled as exact OPT. |
|
|