Launched this week
Memora-Swarm
The TLS for autonomous agent state.
15 followers
The TLS for autonomous agent state.
15 followers
The AI giants run agent swarms on last-write-wins databases like Redis. Result? Silent state forks and swarm-wide memory poisoning (OWASP ASI06). You can't patch this with middleware. Memora replaces fragile DBs with a 3-layer mathematical engine: L1 CRDTs guarantee fork-free convergence. L3 ACFA uses Q16.16 fixed-point math to auto-ban equivocators without a central coordinator. L2 Epistemic logic detects when agents agree on a lie and forces human escalation.
Replacing Redis with CRDTs feels like the right call for agent swarms, and the epistemic layer catching collective hallucinations is genuinely clever. Wish I'd had this last month when two of my agents kept overwriting each other in production.
The CRDT layer sounds rock solid, congrats on shipping this. One thing worth adding is a simple visual replay tool that lets you scrub through the convergence history and see exactly when an agent equivocated or a poisoned value got rejected, since right now debugging a fork in production would still feel like guesswork.
the L3 equivocator banning thing actually worked when I spun up two agents sending conflicting claims, and it just resolved it without me touching anything. Honestly the epistemic layer is what sold me, finally something that flags when my swarm has collectively agreed on nonsense instead of silently shipping it.
The ACFA auto-ban layer sounds solid, but have you considered exposing a lightweight replay mode so teams can audit exactly when and why an agent was flagged as an equivocator? Right now it feels like trust decisions happen in a black box, and being able to step through the Q16.16 evidence trail would make debugging poisoning incidents way less painful for ops folks.
honestly the epistemic lie detection layer sounds promising but i'd love to see a built in replay debugger so we can step through exactly when an agent went rogue before the auto ban kicked in. kind of hard to trust the swarms reasoning if you cant audit the chain after the fact. would also help when explaining it to security teams during reviews
The ACFA auto-ban layer is genuinely clever, finally something that addresses equivocation without leaning on a coordinator bottleneck. Setup was heavier than I expected but once the three layers clicked the convergence behavior felt rock solid.