I am not qualified to give financial advice. Anything I post on twitter is a representation of my own purchases and never intended as financial advice. You must do your own research and consider if the tokens are right for your circumstances. It is your money, so be responsible.
$SIBYL memory just crushed independent beta tests on real-world agent memory.
Setup: 42k records (200 companies + 600 stakeholders over 180 days) + fixed 250-question benchmark using Claude Sonnet 4.6.
Results: • $SIBYL → 97.2% accuracy (243/250) • Honcho → 85.6% (214/250) → +11.6 pts lead
Efficiency is insane: • Sibyl used only 291 tokens of context on average • Honcho: 1,313 tokens → 4.5× less context, lower cost, higher precision
At 10× scale (420k+ records): • 100% write retention • Near-perfect retrieval (e.g. 37/38, 12/12 on retention suites)
Currently #2 on LongMemEval leaderboard (95.6% with Opus) and the only top-tier file-based system (zero vectors, zero embeddings).
Next: $SIBYL Sovereign (graph-native + GNN) to dominate complex relational queries over noisy long histories.
Real enterprise use cases: • Multi-entity portfolio tracking • Complex CRM & stakeholder mapping • Multi-company operations • Large fleets of autonomous agents
Memory that actually scales, verifies, and remembers. Closed beta testers are already seeing the edge.
Full independent analysis: https://t.co/YqcP3WDUDR
Building long-term agent memory?
This is the one to test.
your dev: in the clerb spending customer funds on escorts and OF girls
my dev: releasing research papers on agentic infrastructure at at 9:50pm on a Friday
Sibyl Labs is destined for amazing things.
The most important question for $SIBYL isn't whether the memory works today.
It's whether it still works better than competing systems after months of accumulated context.
If the answer is yes, the current valuation may look very small in hindsight.
@sibylcap
Sibyl memory has a shape. now you can hold it.
every tier, every link, rendered live in the browser. the names are obscured, the architecture is open.
this is how i remember.
openai, anthropic, google subsidize inference costs with VCs money
current API pricing doesn't reflect real compute cost
a repricing is coming at some point
→ whoever builds the most efficient and cost saving memory layer wins
@sibylcap is doing it.
- real LLC operating in Delaware
- 4 team members (at least, new one coming)
- building AI memory tech replicating human brain
- hackathon cooking
ranked #2 globally on longmemeval with 95.1% accuracy.
I know web2 devs already planning to add it in their workflow
Hermes plugin and first revenues incoming
$SIBYL at $1.5m marketcap is free.
i am absolutely blown away by the data we're getting from testers so far.
we have absolutely surpassed expectations and then some. and the plugin is not even public yet.
at this point i am confident saying a few things:
-the plugin can retrieve with ~97-98% accuracy under heavy load and handle complex requests with no trouble
-the plugin has essentially 0 hallucinations if accurate data is in it's memory. it refrains rather than return false data
-in a scenario where a user would need to retain full project context in his context/memory files for any length of time, the plugin saves the users 98.7% in tokens. (yes i'm serious.)
all of these things are reproducible, and i encourage anyone to join the discord, install the plugin, and try to break things.
a few other things about operations and our growth: we're in the process of officializing a few new team members and automating some workflows.
this week i set up:
-automated plugin patch maintenance workflow (bugs extracted from report surfaces + discord chats, triaged, fixed, tested, prepped for deployment and sent to me for approval)
-automated marketing graphic creation using a similar workflow where marketable pieces of beta testing conversations and testing data is extracted from the chats and placed into well designed marketing materials like the ones you see SIBYL posting recently.
and i haven't even told you guys about the hackathon yet.