atmem/atmem-2.3.10b1-dolphinbench

AtMem 2.3.10b1 — DolphinBench, 600 tasks

AtMem with Hermes and GPT-5.6 Luna passed 390 of 600 tasks (65.0%) on the public DolphinBench release. This dataset provides all task outcomes, the full submission ZIP, recorded usage and timings, and the local validator result.

Persona	Passed	Failed	Success rate
Morgan	135 / 200	65	67.5%
Alex	123 / 200	77	61.5%
Riley	132 / 200	68	66.0%
Overall	390 / 600	210	65.0%

Median recorded task latency: 85.78 seconds. All required checks must pass for a task to pass. The original upstream validator replays the saved grading responses locally; it reproduces the 390 passes without another model call.

Explore and download

• The tasks dataset viewer contains 600 rows: task request, final response, task/check outcomes, latency, usage and an evidence-recovery note where relevant.
• submission.zip: exactly ingestion.json and tests.json, including all 13,539 chronological history records, all 600 conversations, tool traces and recorded grading responses. Approximately 44 MiB compressed; below the website's 256 MiB upload limit.
• results.json: configuration, scores, accounting and provenance.
• validation.json: complete upstream validation summary.
• SHA256SUMS: download integrity checks.

from datasets import load_dataset

tasks = load_dataset("atmem/atmem-2.3.10b1-dolphinbench", "tasks", split="test")
failures = tasks.filter(lambda row: not row["passed"])
print(tasks[0])

The complete conversations are in the ZIP; the viewer is a convenient task-level index, not a replacement for those execution records. These are benchmark evaluation records, not memory-decision training labels.

Configuration

• Memory: published AtMem 2.3.10b1, with separate, frozen memory per persona; local source-episode ingestion; source-coherent lexical/semantic retrieval and agent-requested evidence discovery. No paid storage classifier or source-state assessor in this condition.
• Reader: Hermes, openai/gpt-5.6-luna, high reasoning, 60-turn limit, 1,050,000-token configured context; subscription/OAuth transport.
• Judge: gpt-5.6-sol, medium reasoning, 8,000-token cap; OpenRouter with Azure-only routing, rather than a direct Azure endpoint.
• Benchmark: mem0ai/dolphinbench, revision 81cb6f8405b40a9e76089cef650806a80af06ea2.
• Run: October 22–23, 2026. Fresh conversation and app state per test; original grading specifications and fixed memory checkpoints preserved.

Cost and usage

Recorded ingestion/execution incremental API cash is $0: ingestion is local and the reader consumes subscription capacity. This is not a zero-cost system: subscription fees, local compute and energy are not included or priced here. The reader's recorded usage is retained in the ZIP and validator summary.

Paid judging recorded $3.346938 of completed charges, separately from reader cost. An additional $0.276770 timeout reservation remains uncertain and is not relabeled as a zero charge. One failed subscription request has unknown usage; reported token totals therefore do not include that unknown usage.

Provenance and review status

This is a self-tested public-benchmark result, not an independently verified leaderboard result. Earlier public-task subsets informed system development; this campaign is not an untouched holdout. No matched Mem0 or other comparator was rerun in this campaign.

The run reused the original chronological source-ingestion state from an earlier AtMem artifact. Its original receipts, timings and checkpoints were retained; compatibility and identical frozen state were checked against installed 2.3.10b1. The 600 test interactions were newly executed for this condition.

For Morgan/129, a failed restart truncated the separate backend log. The complete native conversation was recovered without another reader call, and the official transcript-based grading path was used. That row is marked in the viewer; the independent backend log is not claimed to have been recovered. Its native runtime duration excludes outer adapter overhead. Other operational recoveries and byte-identical storage relocations are summarized in results.json.

The package uses upstream clean_message to omit provider-only metadata from copies of recorded judge messages. Original verdicts, response content, token usage and timings are unchanged; raw records remain in the private audit archive. Acceptance of the transport and ingestion provenance is for DolphinBench's maintainers to determine.

Verify locally

Use the pinned upstream repository and its participant requirements:

git clone https://github.com/mem0ai/dolphinbench.git
cd dolphinbench
git checkout 81cb6f8405b40a9e76089cef650806a80af06ea2
python -m pip install -r requirements.txt
python -m harness.export_submission --release . --check /path/to/submission.zip

This validates saved evidence and grading without model calls. Submission ZIP SHA-256: 78788ad3455d04407c037706d3ff129d56eafff89beaac101f2a461cb58d7538.

Attribution

DolphinBench is developed by Mem0. Source histories, tasks, simulated apps and grading are from the pinned DolphinBench repository, distributed under Apache-2.0. This release adds AtMem execution records and the task-level result index. See LICENSE. It contains simulated benchmark personas, not AtMem customers' production conversations.

Stable release promotion

The tested build 2.3.10b1 was promoted to stable AtMem 2.3.10. The qualification checked all 370 shipped product files: only two explicit version constants changed; retrieval, storage, governance, prompts, defaults, dependencies and companion/bridge implementation remain unchanged. Qualification; publication checks.

This remains the original 2.3.10b1, 390/600 recorded run, not a new stable-version benchmark. Original task records, ZIP and evidence revision are unchanged.