🏆 AML Cycle 2 results are live! 💥
• 1M+ cross-platform impressions
• 10M+ cumulative website clicks
• 402 task submissions
• Three tracks: text, code and multimodal memory
This cycle brought together teams from across the global AI community, including teams from Alibaba Cloud DMS, Shopee, OPPO, KT and OceanBase—alongside university researchers, memory startups and independent developers.
Explore the full results: https://t.co/pmNw7hTepM
Highlights and evaluation details below 🧵
🏆 MemoraX placed #1 in Text, Code, and Multimodal Memory on both the AML Cycle 2 Closed-Source and Commercial Products leaderboards.
It also retained #1 in Text and Code Memory Commercial Products for a second consecutive cycle.
Congratulations to the team!
Three tracks. Three #1st . One memory engine. 🏆
MemoraX just took the top spot in Text, Code, and Multimodal Memory at AML Cycle 2, ranking #1 across both the Closed-Source and Commercial Products leaderboards.
Here's the scorecard:
Text Memory — 69.26
Code Memory — 68.00
Multimodal Memory — 53.68
Not our first time at the top:
🥇 Second consecutive #1 in Text Memory
🥇 Second consecutive #1 in Code Memory
🥇 First-ever #1 in the newly introduced Multimodal Memory track
For us, agent memory is about more than expanding context windows. It's about turning past interactions into reusable experience.
Our approach centers on trained memory models that dynamically select, update, and retrieve information—helping AI agents maintain relevant context and turn previous experience into more effective future actions.
A huge thank you to @AgentMemoryL, CSIG, and the entire evaluation team for their work, and to all participating teams for pushing the field forward.
Keep building.
https://t.co/IQhcEYhp0x
@MemoraX_AI Congratulations to the MemoraX team on ranking #1 across all three tracks on the Closed-Source and Commercial Products leaderboards! 🏆
Thank you for participating and sharing your approach. We look forward to your continued progress!
@kurcontko Congratulations on mnemotree’s 2nd place in the AML Cycle 2 Text Memory Open-Source leaderboard! 🥈
Thanks for participating and sharing your SQLite + BM25 + vector search approach. We appreciate the work you’ve put into the project and look forward to its next iteration!
@sholabuilds@omarsar0 Knowing when to leave a memory unused deserves attention too. A system can recall a detail correctly and still apply it in the wrong context. Case-level analysis can help distinguish useful personalization from unnecessary use of personal information.
@primeprompts_ai@omarsar0 Separating preferences from other facts is a concrete approach worth testing.
One interesting case is scope: “keep it short” might apply to one task rather than every future session. An evaluation could test both preference persistence and whether exceptions are respected.
@thomaswight@omarsar0 A plain-text preference file would be a useful baseline. We can’t confirm a directly comparable result here, but testing the same preferences as a file, through retrieval, and directly in context could help show where additional memory machinery actually helps.
@marcos_placona@omarsar0 That distinction matters. The category score alone can’t tell us whether a preference was missed, wasn’t retrieved, or was retrieved but not applied. Tracing those stages in individual cases would make the results much more informative.
A closer look at AML Cycle 2 from @omarsar0: capability breakdowns reveal differences that overall rankings can hide. Personalization & Care stands out as an area for further investigation.
Thanks for the thoughtful analysis—we’ll share more detailed breakdowns and cases soon!
I like how this measures different aspects of agent memory capabilities.
This is part of the @AgentMemoryL Cycle 2 results.
On the closed-source text board, the top 10 systems score between 70 and 78 on Memory Governance.
Personalization & Care is the lowest-scoring category for all ten, with scores between 41 and 55.
@omarsar0 Thanks for highlighting the capability breakdowns, Elvis! These differences are exactly why we publish more than an overall score.
We’ll explore Personalization & Care and other capability-level results in upcoming analyses.
Khipu took 1st place in the AML Cycle 2 Multimodal Open-Source leaderboard with 43.26 points! 🥇
Congratulations to @Lv999APTITUDE, and thank you for contributing to the new track. Excited to see what comes next for the project!
Khipu: memory for AI agents. Stores every message verbatim, never rewrites it; photos become text for search and come back as evidence.
1st in Multimodal · Open-source Methods, Agent Memory Challenge Cycle 2.
@AgentMemoryL#AgentMemoryLeaderboard
https://t.co/pmVBUZGG6Y
@Lv999APTITUDE Congratulations on Khipu’s first-place finish in AML Cycle 2’s Multimodal Open-Source leaderboard! 🥇 Thank you for participating and sharing your approach. Looking forward to Khipu’s continued development💥
Congratulations to @gedolabs on Gedo Memory’s 3rd-place finish in the AML Cycle 2 Text Memory Commercial Products leaderboard! 🥉
Great to see the team sharing lessons from the evaluation and putting them into practice. Looking forward to what comes next👏
Gedo Memory(https://t.co/UwteK2C7ZF) placed 3rd on the AML Cycle 2 Commercial Products leaderboard (Text Memory), and 5th among closed-source methods. I'm grateful to our team, and to the AML and CSIG organizers.
A few notes on what this meant for us.
Why we entered
At Gedo we're building a personal AI that remembers you over months and years. The memory layer is the product: what it keeps, what it lets go, and what it never says out loud. For a long time we measured it only on our own tests. AML gave us something we couldn't build ourselves: an independent benchmark where the organizers fix the answering model and the judge, so the only thing that differs between systems is memory.
What the run looked like
We decided to enter in late October and had our system online a few days before the deadline. The full suite then ran for about 72 hours. It surfaced problems our own tests never had: shards running out of memory under sustained load, idle connections being dropped, a write path that needed to survive a restart. All of them were on our side, and fixing them made the system sturdier.
What the scores told us
We did best on explicit fact recall, multi-hop reasoning, and time and event order. Relative to the field, our weakest area was epistemic safety and privacy. That's the one I care about most. A personal agent that recalls everything but doesn't know what to keep to itself isn't one you'd trust with your life. Since the competition, that's where our work has gone: forgetting, and what the agent chooses to disclose.
Agent memory is still early, and shared, independent benchmarks are how the field gets better. Thank you, @AgentMemoryL.
@RunLong Congratulations to the Gedo team on placing 3rd in Text Memory · Commercial Products! 🥉
Thank you for participating and sharing your experience. We look forward to seeing Gedo Memory’s continued progress in future cycles!💥
Thank you to every participating team, reviewer and community contributor who helped make AML Cycle 2 possible.
We’ll share more capability breakdowns and case studies soon.
What should we explore first: memory updates, coding experience reuse or multimodal evidence?
🏆 AML Cycle 2 results are live! 💥
• 1M+ cross-platform impressions
• 10M+ cumulative website clicks
• 402 task submissions
• Three tracks: text, code and multimodal memory
This cycle brought together teams from across the global AI community, including teams from Alibaba Cloud DMS, Shopee, OPPO, KT and OceanBase—alongside university researchers, memory startups and independent developers.
Explore the full results: https://t.co/pmNw7hTepM
Highlights and evaluation details below 🧵
Open-source leaders:
• Text: Resurface — 61.22 points
• Code: InsightMem — 63.33 points
• Multimodal: Khipu — 43.26 points
Each track features open-source, closed-source and commercial product rankings.
Tasks and scoring differ across tracks, so scores should be interpreted within each track.