The first benchmark for multimodal, multi-source personalized referential memory QA over 3.3 GB of raw multimodal data spanning ~4 years, with evidence-grounded retrieval and answering.
Graphon is #1 and #2 on the leaderboard.
expand ⤢
ATM-Bench / memory systems↗ better
QA score (%)Recall@10 (%)
Best run per system, from the public ATM-Bench leaderboard. The models behind each run and the full results are on the leaderboard.GraphonOthersSource: atmbench.github.io/leaderboard ↗
ATM-Bench’s 31 hardest questions. Each needs many pieces of evidence pulled together across years of emails, photos and videos. In the paper that introduced it, no system reached 20%.
Graphon scores 65.6 with 66.6% Recall@10, ahead of every memory system and coding agent on the board. The best coding agent, Antigravity with Gemini 3.7 Flash, scores 62.9.
All runs
expand ⤢
ATM-Bench Hard / memory systems & agents↗ better
QA score (%)Recall@10 (%)
Best run per system, from the public ATM-Bench leaderboard’s Hard track. Graphon’s run is submitted to the leaderboard and awaiting review. Coding agents report no recall, so they stand in the band on the left, placed by QA score; the best run of each is shown.GraphonOthersCoding agentsOur submission: github.com/atmbench/atmbench.github.io/pull/9 ↗
ATM-Bench Hard: the best run of each system and coding agent
System
Type · model
QA score
Recall@10
Graphon V1
Memory Graphon
65.6
66.6
Memexa v2
Memory DeepSeek V4.1
63.6
63.0
Antigravity
Coding agent Gemini 3.7 Flash (high)
62.9
–
Claude Code
Coding agent Claude Fable 5.1 (high)
61.0
–
Codex
Coding agent GPT-5.6 Sol (medium)
58.8
–
Kimi Code
Coding agent Kimi K3-256k (high)
52.5
–
OpenCode
Coding agent Qwen3.8-Flash-Next (xhigh)
51.7
–
Pi
Coding agent Qwen3.8-Flash-Next (xhigh)
51.6
–
OpenClaw
Coding agent Kimi K2.5
25.4
–
A-Mem
Memory DeepSeek-V4-flash
16.7
38.5
ATM-RAG
RAG Qwen3-VL-8B-Instruct
13.8
30.4
MemoryOS
Memory Qwen3-VL-8B-Instruct
13.7
32.7
MemPalace
Memory DeepSeek-V4-flash
13.6
28.9
HippoRAG2
RAG DeepSeek-V4-flash
12.3
32.2
Mem0
Memory Qwen3-VL-8B-Instruct
9.2
23.7
SimpleMem
Memory Qwen3-VL-8B-Instruct
3.2
7.0
Best run per system or agent harness, from the public ATM-Bench leaderboard’s Hard track. Coding agents report no Recall@10. Oracle rows, which give the model the gold evidence, are left out.
LoCoMo
A benchmark for very long-term conversational memory: dialogues of about 600 turns across up to 32 sessions, with questions about what was said, when, and how it connects.
Graphon redraws the cost-accuracy Pareto frontier: 94.0% accuracy at 7.9K tokens per question, more accurate than full context with less than half the tokens. Every answer (JSON) ↓
expand ⤢
↖ betterLoCoMo / accuracy vs. cost
Judge accuracy (%)Tokens per question (log) · cheaper ←
Baseline results and the published frontier are as reported in independent third-party evaluations. Graphon redraws it: full context drops off, and the frontier now runs through Mem0, PropMem and MemCurrent up to Graphon. Systems in the shaded region right of and below Graphon V1 are Pareto-dominated by it.GraphonOthersNew frontierPublished frontierReplaced
ConvoMem
A benchmark of 75,336 question-answer pairs over long conversation histories, testing user facts, assistant recall, preferences, facts that change, implicit connections, and knowing when not to answer.
Graphon pushes the cost-accuracy Pareto frontier up: 91.0% accuracy at 8.9K tokens per question, more accurate than full context with a twentieth of the tokens. Every answer (JSON) ↓
expand ⤢
↖ betterConvoMem / accuracy vs. cost
Judge accuracy (%)Tokens per question (log) · cheaper ←
Baseline results and the published frontier are as reported in independent third-party evaluations. Every system on it stays on the frontier; Graphon extends it upward. Systems in the shaded region right of and below Graphon V1 are Pareto-dominated by it.GraphonOthersExtended by GraphonPublished frontier
LongMemEval
A benchmark of 500 questions over long chat histories of about 50 sessions each, testing recall of what the user and the assistant said, reasoning across sessions and over time, facts that change, and preferences.
Graphon pushes the cost-accuracy Pareto frontier up: 92.8% accuracy at 15.1K tokens per question, more accurate than full context with about a seventh of the tokens.
expand ⤢
↖ betterLongMemEval / accuracy vs. cost
Judge accuracy (%)Tokens per question (log) · cheaper ←
Baseline results and the published frontier are as reported in independent third-party evaluations. Every system on it stays on the frontier; Graphon extends it upward. Systems in the shaded region right of and below Graphon V1 are Pareto-dominated by it.GraphonOthersExtended by GraphonPublished frontier
Maximum accuracy
The frontier above is Graphon tuned for efficiency. Graphon also offers modes with deeper graph reasoning, which spend more graph traversals for higher accuracy. In that mode it answers 481 of the 500 questions correctly (96.2%), the highest published score.
LongMemEval: the highest published accuracy of each system
Accuracy is the share of the 500 questions an LLM judge grades correct with LongMemEval’s grading prompts. Retrieval scores such as recall are not counted. Graphon’s file lists, for every question, the question, the dataset’s answer, Graphon’s answer, and the judge’s verdict and reason.
A benchmark of 2,000 questions over conversations of up to 10 million tokens, testing ten memory abilities, including instructions and preferences, facts that change, contradictions, and the order of events.
At 10 million tokens, Graphon scores 73.0, ahead of Exabase M-1 (68.0), Mastra (66.4) and every other published result under BEAM’s standard evaluation. Every answer (JSON) ↓
All tiers
expand ⤢
BEAM-10M / overall score→ better
Overall score, all 10 abilities (%)
Baseline results are as reported by each provider. Graphon’s answers are graded with BEAM’s rubric prompts by Gemini 3 Flash.GraphonOthers
A benchmark for agent memory in action: three simulated knowledge workers with years of history each, and 600 tool-using tasks that pass only if the agent recalls and applies the right fact. Every run reports its total cost and latency alongside accuracy.
Graphon passes 504 of 600 tasks (84.0%) for $11.84 in total, with a median of 28.8 s a task. The best run on the public leaderboard, on the same model, passes 424 (70.7%) for $96.21.
That $11.84 is the whole bill for the run: Graphon’s indexing and search, plus every input and output token GPT 5.6 Luna spent working through the 600 tasks.
Graphon’s run (Graphon + GPT 5.6 Luna) is our own. Every other run is from the public DolphinBench leaderboard. Hover a point for its setup, cost and latency.GraphonHermes + GPT 5.6 LunaHermes + MiniMax M3Claude Code + Sonnet 5New frontierPrevious frontierOur submission: github.com/mem0ai/dolphinbench/pull/12 ↗
DolphinBench: every run
Memory
Harness + model
Passed
Accuracy
Total cost
Median latency
Trusted by
At the 49ers, we're always looking for innovative ways to strengthen our business and drive growth. From uncovering new sponsorship opportunities to empowering our teams, we're excited to explore how Graphon Labs' technology can help us gain a competitive edge.
Costa KladianosEVP of Technology, San Francisco 49ers
At Hitachi Rail, we seek to unlock greater value from complex and distributed information landscapes, often with mission-critical operational outcomes. Our evaluation of Graphon Labs' innovative technologies shows promise to facilitate the discovery of context knowledge, automate the connection to data sources and improve AI inferencing results in our specific industrial context.
Amaury JourdanDeputy Manager and CTO for Digital and AI Activities, Hitachi Rail
S3 compatible fast, reliable, and secure file storage.
Search
Ultra fast keyword and semantic search over markdown, json, pdf, images, videos. Zero config. Replaces elastic search in your dev stack
Answer
Natural language question and answer API for humans and agents with source citations. Zero config, SOTA accuracy and speed.
Research
Powerful deep research question and answer for complex research and analysis across TB of data.
how it works
From points to a shape.
the context problem
Gigabytes of multimodal data.
Video, audio, PDFs, text, markdown, JSON and spreadsheets. One square is one file: 1,024 here, millions in practice.
the usual way
Vector search sees points.
It cuts your data into chunks, turns each one into a point, and returns the nearest few. Everything between the points (the reasons, the sequence, the relations) is gone.
12 points. the rest? gone.
relate
Graphon reads it all at once.
Graphon builds one large graph of your data, then fits a graphon to it: the limit object of that graph, a single function that maps the topology of the full graph.
global structure
across every file, not chunk by chunk
inferred schema
learned from the data, no ontology to draw
scope
the whole corpus
strong linkweak linkacross topics
there’s always one that’s late →
remember
Graphon learns the shape between them.
A graphon is the continuous limit of a graph: one smooth function over all of your data instead of a pile of points. Nothing between the points is lost.
← this is the good part
retrieve
Ask, and the graphon guides the search.
A question first finds a few matches. From there the graphon steers the search toward what is connected to them, until it has every source the answer needs.
How do we set the charger voltage on the APS-200 auxiliary converter?
manual.pdf · p. 47training.mp4 · 12:40ticket #4471
The manual says 110 V, set in menu 4.2. The training video adds a step: isolate the 750 V supply and wait five minutes first. A ticket notes units past serial 2400 read 2 V high, so set those to 108 V.
