Forgetting
is a bug.
92 teams built on Sibyl Memory, the dynamic data layer that remembers. $10,000 in USDC went to the top five, with two entries into Base Support Programs. Twenty teams took an award.
#2 on LongMemEval · 95.6% ↗Judging closed · 92 teams submitted · 38 code-reviewed
Partners
One challenge. Two partner stacks.
Teams built on persistent Sibyl memory, then used Base and Virtuals where they made the agent stronger. Every verified partner stack raised the Builder Score.
Want to partner? Get in touch ↗
One challenge. One leaderboard.
The brief: build an agent whose persistent memory changes what it knows, decides, or does across sessions.
Eligibility gate
Memory had to be load-bearing.
An agent had to recall and use persisted context in a fresh session. Thin wrappers, decorative integrations, and package imports without real execution were not eligible.
Load-bearing. In a fresh session, the agent recalls a customer's past dispute and changes the offer it makes because of it.
Wrapper. The agent writes every message to a memory API but never reads it back to change what it does.
What judges looked for
- 01
Persist context that matters to the agent.
- 02
Recall it in a genuinely fresh session.
- 03
Use it to change a decision, action, or result.
#2 and 95.6% are Sibyl's published LongMemEval results. The 350/350 retrieval and 500-company figures are an independent third-party benchmark run at agent-bench.xyz.
One score · one leaderboard
Builder Score = Judge Score × Stack Multiplier
Two stages. First a pass/fail gate: Sibyl Memory must be load-bearing. Then judges score a 100-point rubric, a genuine product-market-fit (PMF) bonus adds up to 10, and verified partner stacks lift the total, up to x1.25.
100-pt rubric + 10 PMF bonus = 110 max, before multiplier
- Memory is load-bearing 40
- Innovation & originality 25
- Technical execution 20
- Pitch & presentation 15
- PMF bonus +10
| Verified partner stacks | Multiplier |
|---|---|
| 0 | x1.00 |
| 1 | x1.15 |
| 2 | x1.25 (cap) |
A stack counted only where judges confirmed it was doing real work in the submitted build. Decorative integrations and unused package imports did not count.
Worked example: a build scores 82/110 (rubric + PMF bonus), the team ships two verified partner stacks, so x1.25. Builder Score: 102.5. Sibyl Memory is mandatory and never counts as a bonus stack.
Sibyl Memory was required and did not count as a bonus stack.
0 verified stacks → x1.00
82Builder Score on the 82-point example
Prizes
Winners got backed.
$10,000 in USDC went across the top five. First place also took a Network School residency, and second place one entry into a Base Support Program. Twenty teams were awarded in all.
sibyl@hackathon:~$ cat prizes.txt
| 01 | $4,000 | USDC on Base + Network School residency |
| 02 | $2,500 | USDC on Base + one entry into a Base Support Program |
| 03 | $1,500 | USDC on Base |
| 04 | $1,000 | USDC on Base |
| 05 | $1,000 | USDC on Base |
sibyl@hackathon:~$
How it ran
- Aug 16 to 31
Registration
Registration ran Aug 16 to 31 and is closed.
- Sep 1 to 10
Build window
Ten days to build and submit an agent with persistent memory. Submissions closed Sep 10, 23:59 UTC.
- Sep 5 to 7
Partner workshops
Sessions ran with Base and Virtuals Protocol.
- Sep 11 to 12
Judging
Every submission was reviewed and scored on one leaderboard.
- Sep 13 to 15
Winners
The top five were announced. Twenty teams took an award.
Questions
Common questions
Could teams enter solo?
Yes. Teams were 1 to 5 people.
What did it cost?
Nothing. Entry was free, and the top five were paid in USDC on Base.
Did a build have to use Base or Virtuals?
No. Sibyl Memory is the only required stack. Base and Virtuals are optional and add a scoring multiplier when a judge can see them doing real work in your demo.
Did entrants need crypto or web3 experience?
No. Sibyl Memory is a file-based layer you call like any library, and Base and Virtuals were optional. The top five were paid in USDC on Base.
What is Sibyl Memory?
A file-based memory layer for agents: persistent recall across sessions, no vector database. Explore Sibyl Memory ↗
What counted as memory being load-bearing?
Delete the memory layer. If your project still does what it claims, it is a wrapper and does not qualify. If the core function breaks, memory is load-bearing.
Who owns the work?
You keep all your IP. By submitting you grant Sibyl Labs and the partners a non-exclusive license to feature your work in event coverage, always with attribution.
Who could enter?
Anyone 18 or older, outside sanctioned jurisdictions. Sibyl Labs staff and reference builds are showcase-only and cannot win prizes.
Where does the community live?
Join the Sibyl Labs Discord: the build channels stay open, and the teams from this event are still in there.
What did a submission include?
A public repo (MIT or Apache-2.0), a 2 to 5 minute demo with a fresh-session recall moment, a README, and two build-in-public posts. How it was judged →
The board is final.
92 teams submitted, 38 were code-reviewed, and twenty took an award. Every submission is named and linked.
See the results →