Loading
Loading
Most memory-benchmark numbers come from single-vendor pipelines — each system scored against its own reader and its own judge, so the numbers don't compare. We held everything fixed except the memory system: LongMemEval-S, n = 500, one shared Qwen 3.5-9B reader, and four independent judges — including mem0's own verbatim judge prompt. Atlaso wins on every one.
The four-judge structure is the point. One judge can favor a system's answer style; four independent prompts — one of them mem0's own — give a sensitivity band instead of a single number. Atlaso wins in every band, from +9.8pp to +14.8pp.
LongMemEval-S · n = 500 · identical question IDs across arms · shared reader: Qwen 3.5-9B · four independent judges (Anthropic Haiku 4.5 strict, mem0's verbatim judge prompt, OpenAI GPT-5 permissive, OpenAI GPT-4o strict). Answers are label-blind normalized so a judge can't tell which system produced them.
Memory benchmarks are dominated by single-vendor pipelines: each system is scored end-to-end against its own generator, its own reader, and its own judge. Line two of those numbers up in a deck and the comparison can be off by fifty points purely because the two were generated under different rules.
mem0 publishes a headline 93.4% on LongMemEval-S. Running mem0's default OSS pipeline on the same 500 questions, scored with mem0's ownverbatim judge prompt, we observe 44.2% — a 49.2pp gap under mem0's own rule. We did not reproduce mem0's full managed pipeline, so the honest reading is “methodology + pipeline,” not methodology alone. What is certain: under matched conditions, Atlaso leads on the same fixture and the same judge.
mem0 published — own pipeline, own judge
mem0 default — matched conditions, mem0's own judge
Atlaso — same conditions, same judge
The honest comparison floor sits somewhere between mem0's 44.2% and their published 93.4%; work that reproduces mem0's full managed-platform pipeline will narrow that band. We publish the whole reasoning in the study.
LoCoMo adversarial subset · n = 200 · Haiku 4.5 strict
mem0-default
Atlaso (−11.5pp)
On LoCoMo's temporal-reasoning and planted-distractor questions Atlaso loses to mem0 by 11.5 points. The substrate is currently tuned for questions whose answer either clearly exists in the field or does not; against adversarial distractors it over-abstains. We report the loss because where a system fails is more diagnostic than where it succeeds.
mem0's default configuration runs a gpt-4o-miniextraction call on ingestion — an LLM summarizes each turn before it's stored. Atlaso writes its deposits with zero LLM callsat ingestion. That is a genuine architectural difference, not a scoreboard claim: it changes ingestion cost and behavior, and it's worth knowing which one you're buying.
Read the full study — methodology, run logs, and the limits we publish
They solve related problems at different layers, so the right pick depends on what you're building.
mem0
An open-source memory SDK
mem0 is a library you wire into your own application — you own the storage, the extraction configuration, and the retrieval calls. If you're building a product and want memory as a component you control end-to-end, that's the shape it fits.
Atlaso
A hosted memory layer for the tools you already use
Atlaso connects to the AI tools you already work in and captures and recalls memory automatically — no app to build, nothing to wire up. It works across:
Free for one device and one tool with unlimited memories; Pro is $10/month for unlimited tools and devices sharing one memory.
Connect Atlaso to the tools you already use and stop starting from zero. Free to start — no credit card required.