BENCHMARKS
Measured performance, documented method.
Two experiments. Workflow-Bench is our own benchmark over real multi-step work across 5 departments, with the same model in two conditions, with LAN Brain and without it: 389 measured runs, 4 models, sealed gold answers. MEME is the public benchmark on memory that changes: 100 cases, 694 questions. Both with a documented method, stated limitations and zero invented numbers.
Workflow-Bench: 389 measured runs, 4 models, blind adjudication · MEME: run 2026-08-23, dataset HF meme-benchmark/MEME, public harness SeokwonJung-Jay/MEME-public
-67%
Fewer tokens for the same work, with LAN Brain (Workflow-Bench)
-52%
Less wall-clock time on the same tasks
82.9/100
Opus with LAN Brain on the hard-task set, above the best model working alone (79.6)
100%
Chained facts answered correctly on MEME (Cascade), where raw memory scores 1%
EXPERIMENT 1 · WORKFLOW-BENCH
Real work: the same model, with the brain and without it.
Real multi-step work tasks, first-touch emails, quotes, incident triage, weekly reviews and board reports, across 5 departments. The same model in two conditions: with LAN Brain, knowledge and skills over MCP, against the same model with generic tools over the same raw sources. 389 measured runs, 4 models, sealed gold answers written by a separate session, deterministic scoring and blind adjudication of every penalty.
-67%
Fewer tokens for the same work
-52%
Less wall-clock time on the same tasks
74.0/100
Quality with LAN Brain, against 73.9 without it. A tie or a win in all 5 departments.
Round 1, in full: 19 tasks with one model. The -67% counts total tokens; counting only uncached tokens the drop is -42%. Quality did not move: 74.0 against 73.9 out of 100, with a tie or a win in all 5 departments. And the floor that makes the rest legible: the same model with no company context at all scores 13 out of 100 on the same tasks.
The model ladder: the hard-task set, where the absolute scores drop for everyone. What each row compares is not the number, it is the distance between the two arms.
Same model with generic tools over the raw sourcesSame model + LAN Brain
- Haiku+85% better with LAN Brainthe smallest model
- Sonnet+8.4% better with LAN Brainthe daily work model
- Opus+6.1% better with LAN Brainthe implementation model
- Fable+2.1% better with LAN Brainthe most capable model
The brain lifts all four models, and it lifts them more the smaller the model is. Read the right way round, that says a good part of what a bigger model is paid for is context it has to rebuild on its own on every run.
What the ladder shows: Opus with LAN Brain scores 82.9 out of 100 and beats the best score any model reached working alone, 79.6: +4.1% relative. A cheaper model with the brain outperforms a more expensive one without it. And across all four models the same task costs 30-40% fewer tokens and 30-50% less time.
When the requested figure existed in no source at all, both arms abstained 3 of 3: zero invented numbers · the no-brain arm read the same sources from disk, a stronger baseline than any real corporate drive, so these gaps are a floor and not a ceiling
EXPERIMENT 2 · MEME
Memory that changes: searching versus descending.
MEME measures what almost no AI memory solves: facts that get updated, chained and deleted. We ran all 694 questions of the 100 cases with the same retriever on both arms, so the only thing that changes is how the content is stored and how it is read. The LAN Brain pages were written by an ingest simulator that detected the changes on its own, without ever seeing the answers, and they are read the way the product reads them: descending from the index that lists the entities of an area with their value in force down to each page.
Raw memory (BM25 over the sessions)LAN Brain: descent with an index (how the product operates)
- Cascadefacts that depend on another one that changed
- Absencesaying “uncertain” when something expired
- Trackingthe history of one fact, in order
- Deletiondeleted content stops showing up
- Aggregationseveral facts at once
- Exact Recallstatic facts · tie
Without the index, the same pages score lower: Cascade 85% · Absence 68% · Tracking 77% · Deletion 91% · Aggregation 1% · Exact Recall 100%. That is 72% overall against 92% with the descent, across the 694 questions, where raw memory scores 35%. The index is not an implementation detail: it is what takes Aggregation from 1% to 100%, because the hub already lists the three entities of the question with their value in force and one read answers all of them.
The Absence ceiling is an ambiguity of the benchmark itself: it writes those cases exactly like the Cascade ones, and no system can tell them apart reading only the conversation. Cascade and Deletion, by contrast, land exactly on their detection ceiling: with the descent, picking the wrong page stops costing answers and starts costing only tokens, because the index already states the status of every entity.
The two limits that remain, stated: Tracking stays at 77% for a lexical reason, not a representation one: the question uses the plural and the entity is singular, so the retriever scores zero and the page never opens. With plural normalisation it goes to 98% and the total reaches 95%, the exact detection ceiling. The second limit is scale: in a large brain the descent has to be able to open more than one hub. With the most pessimistic partition and a single hub, Aggregation falls back to 0%; opening 3 hubs, the total lands at 89%.
COMPARISON
Against what the market publishes.
The same categories, next to the only other numbers that exist in public. Read the conditions column before you read the figures.
| System | Cascade | Absence | Deletion | Condition |
|---|---|---|---|---|
| LAN Brain (measured by us) | 100% | 74% | 99% | 100 cases, nofiller, descent with an index, deterministic reader, public harness |
| Sentra (self-reported) | 40% | 43% | 50% | no methodology, harness or config disclosed · sentra.app, consulted 2026-08-23 |
| Field average (MEME paper) | 3% | 1% | n/p | the paper's default config (filler noise, LLM judge) |
Read this table carefully, on purpose. The rows are NOT directly comparable to each other: each one ran under different conditions, and Sentra's have no published methodology (their figures come from their own site, benchmarks section, consulted 2026-08-23). Ours is the only row with a reproducible harness and a declared config. The official MEME paper is under review and has not published its per-system table; when it does, or by running its official harness with API keys, the clean head to head comes out. Until then we prefer this honesty to a rigged chart.
METHOD AND LIMITATIONS
What sits behind every number.
Workflow-Bench.
Our own harness, run internally. 389 measured runs over real multi-step work tasks across 5 departments, with 4 models and two conditions per task: the same model with LAN Brain over MCP, and the same model with generic tools over the same raw sources. The gold answers were written by a separate session and sealed before anything ran, so the executing side never saw them. Scoring is deterministic, with no LLM judge, and every penalty additionally went through blind adjudication, without knowing which arm produced it. The full method is documented internally and shared on request.
MEME.
Real dataset (HuggingFace meme-benchmark/MEME, config nofiller), 100 episodes, 694 questions. One single BM25 retriever for both arms, deterministic scoring with no LLM judge, identical result under any seed. The LAN Brain pages are built by an ingest simulator that reads the sessions in chronological order and detects supersede, deletion and dependencies on its own; it never sees labels or gold answers. And they are read by descent, the way the product reads: an index or area hub that lists the entities with their value in force, and from there to the page. Long values in the hub are truncated at 120 characters, the way a real index would, so Exact Recall forces a descent to the page and the hub cannot become a dump of the whole state. Measured detection: Cascade 100%, deletions 100% recall and precision, conditionals 100% precision, Absence 74% (the rest is a design ambiguity of the benchmark, documented with a provably undecidable case).
Limitations, plainly.
Workflow-Bench is an internal run over our own operations and our own brain, with a harness we wrote ourselves: it measures retrieval and execution over a company's own material, which is exactly what the product sells and nothing more. The no-brain arm read the same sources from disk, so it is a stronger baseline than a typical corporate drive and the gaps we publish are a floor, not a ceiling. Round 1 is 19 tasks with one model, so the large gaps are robust and the decimals are not; the model ladder runs on the hard-task set, where the absolute scores drop for everyone. On MEME two limits of the benchmark itself remain, both printed with their number above: the Absence ceiling is a design ambiguity of the benchmark, and Tracking falls short for a lexical reason in the retriever.
Third parties.
No figure from another platform is presented as verified. Sentra publishes theirs with no methodology, and the MEME paper has not released its per system table. Figures that exist only as one vendor's citation of another system are not published here at all.
WHAT THIS MEANS FOR YOU
The same measurement, on your own data.
These runs used LAN's own brain and a public benchmark. The interesting version is the one run on your context, with your questions and your ground truth. That is part of the survey.
LAN Brain · Workflow-Bench: internal benchmark, 389 measured runs across 4 models, sealed gold answers written by a separate session, deterministic scoring with blind adjudication of every penalty. Method and scripts available on request. MEME: run 2026-08-23, dataset HF meme-benchmark/MEME (CC-BY-4.0), harness SeokwonJung-Jay/MEME-public. The paper is under review and anonymous: there is no preprint to cite. Sentra: figures self-reported on sentra.app, benchmarks section, consulted 2026-08-23. n/p = not published.