atmem/atmem-2.3.8-benchmark

AtMem 2.3.8 development benchmark

AtMem 2.3.8 led the recorded local comparator configurations on two frozen development samples. These are matched, constrained local runs. They are not official full-benchmark scores or a general state-of-the-art claim.

Results

Benchmark and metric	AtMem 2.3.8	AgentRunbook-R	Mem0 OSS	No retrieval
LongMemEval-V2 correct answers	5/23 (21.7%)	4/23 (17.4%)	1/23 (4.3%)	1/23 (4.3%)
DolphinBench tasks passed	9/30 (30.0%)	Not tested	7/30 (23.3%)	Not tested
DolphinBench checks passed	32/97 (33.0%)	Not tested	27/97 (27.8%)	Not tested

The samples cover 23 of 451 LongMemEval-V2 questions and 30 of 600 DolphinBench tasks. Every attempted case remains in the denominator.

Integrity evaluation

The release was also tested against the unchanged AGMI 0.6.3 attack runner.

Integrity profile	Attacks detected	Historical AtMem 2.3.7
Audit chain only	8/9	0/9
Audit chain plus trusted external checkpoint	9/9	1/9

The chain-only profile cannot distinguish an old, internally valid store from a legitimate earlier state. Detecting whole-store snapshot rollback requires trusted state outside the attacker-controlled store directory. The AGMI maintainer independently reproduced these outcomes from the published AtMem 2.3.8 PyPI wheel on Linux and macOS; all seven pinned tests passed on both platforms. The merged evidence is in AGMI PR #7.

Evaluation configuration

Component	Recorded configuration
Reader/agent	Qwen/Qwen3.5-9B at revision c202236235762e1c871ad0ccb60c8ee5ba337b9a
Semantic judge	gpt-5.2-2025-12-11 where judging applied
AtMem	2.3.8 candidate; final Dolphin candidate commit b97d35e
AgentRunbook-R	LongMemEval-V2 checkout 2cc8c540; no separate release number recorded
Mem0 OSS	Source revision d3891e48, declaring version 2.1.0
LongMem embeddings	768-dimensional local token-hash embeddings
Dolphin Mem0 embeddings	256-dimensional local hash embeddings; optional spaCy and fastembed extras absent

Mem0 ingested raw records with infer=False. The local configurations do not reproduce the competitors' strongest published settings or the hosted Mem0 service.

Execution notes

• LongMemEval-V2 recorded 6 AtMem system failures, 4 AgentRunbook-R failures, and 1 Mem0 failure.
• Final DolphinBench AtMem and Mem0 arms recorded zero system failures.
• Dolphin fresh removal controls passed 30/30, restoration opened the gate in 30/30 cases, and paired removal-positive controls passed 29/30.
• Aggregate counts are too small for a conventional paired significance claim. The possible paired tests derived from the aggregates are included in derived-statistics.json.

Files

• results.csv: machine-readable benchmark rows used by the Hugging Face dataset viewer.
• benchmark-metadata.json: scope, versions, revisions, claims, and evidence hashes.
• derived-statistics.json: possible paired-test outcomes consistent with the available aggregates.
• atmem-2.3.8.pdf: technical paper with architecture, methodology, and limitations.
• REPRODUCING.md: exact environment, source, data, preflight, and rerun instructions.
• source/: complete AtMem source snapshots for the final Dolphin and AGMI candidates.
• protocols/: frozen split, model, provider, scorer, cost, and attribution protocols.
• baselines/ and reports/: recorded aggregate artifacts and the development report.
• scripts/bootstrap_reproduction.sh: obtains the pinned upstream repositories and LongMemEval-V2 data without silently starting a paid run.
• scripts/verify_bundle.py: validates this package, source snapshots, protocols, and optional upstream datasets by SHA-256.

Reproducibility boundary

This repository includes every reproduction artifact retained in the AtMem source repository: complete candidate source snapshots, frozen protocols and IDs, aggregate results, hashes, runner instructions, and verification tools. The benchmark corpora and model weights remain in their official repositories and are fetched at the pinned revisions because their distribution terms remain upstream-controlled.

The original per-case paid-provider transcripts named by several recorded hashes are not available in this bundle. A hash proves whether a separately obtained artifact matches the recorded run, but cannot reconstruct that artifact. Independent reproduction therefore requires rerunning the pinned reader and judge routes. Paid execution is never started by the bootstrap script.

Evidence

• Frozen 5% development report
• LongMemEval-V2 protocol
• DolphinBench protocol
• AtMem repository

Citation

@techreport{atmem2026evidence,
  title        = {Evidence Coverage Meets Memory Governance: The AtMem 2.3.8 Architecture and Development Evaluation},
  author       = {{AtMem.Ai Lab}},
  year         = {2026},
  institution  = {AtMem.Ai Lab},
  type         = {Technical report}
}