this is f**king insane researchers built a 4-agent coding system that learns from every problem instead of resetting after each run [4 agents, 1 evolving memory] 1. Planner picks the strategy 2. Solver writes the code 3. Oracle verifies the result 4. Hacker tries to break it 5. pass/fail... updates the knowledge graph every agent keeps its own graph-structured memory, so the next problem starts with what the system already learned [no retraining. no weight updates.] across 4 coding benchmarks, Solvita nearly doubled single-pass baseline accuracy save this so you don’t lose it.show more

I found this graph here

if the system can actually carry lessons from failed runs into the next task without retraining, that's a huge shift for coding agents

It’s called Claude.

Bro this is gold!

bro, you're damn right because I just found out about this recently and I love it so much

the real upgrade is learning between runs

Powerful stuff here, sharing with all my friends now

a 4-agent coding system with an evolving memory graph that learns from every bug is the ultimate breakthrough

learning from your mistakes is a very useful feature for agents

cyberpunk detected

this is for real insane bro..

four agents enough to handle all the tasks?

I think so, maybe even 1 agent would be enough

This is really nice

Thanks Avid, by the way, have you seen this yet or not?

Yes 🙌🏻

this is very cool🫡

this is genuinely the future of ai right here

In fact, this graph stores almost all the important notes and articles, and its operating principle is similar to obsidian but it's more like a library

@gippp69 pretty wild. wonder how long til it codes itself out of a job 😂

This is formalised in Dynamic Contradiction Logic as Multi-Mirror Computing. Shared memory is the mutual projection of each agent’s fractal DCL intelligence. The shared Tapestry is the projection space itself.

Agents need experience, not just context

memory turns every solved problem into leverage

this is a real evolution in the agent's work

yes, for agents it's a salvation, to be honest

no way its real bruh

graph memory is the ultimate meta

Every agents knows what to do

The interesting part isn’t the 4 agents. It’s that failures stop becoming expensive resets.

i need system like that

You inspired me to try this

bookmarked gipp, ur loop engineering piece pairs perfect with this

what's the forgetting mechanism? learning from every run is also how it locks in its own bad habits

The system gets better by changing what it remembers, not what it weighs.

Evolving memory makes multi agent coding feel inevitable.

i save it

memory actually compounds

The interesting shift is that the agents aren’t just collaborating. They’re accumulating institutional memory. If the memory stays reliable, every solved bug becomes a better starting point instead of another cold start. That could matter more than adding another agent.

4 agents and 1 memory that's all you need

Where did it mention it's agent-based? This website existed even before LLMs got this hype. I've been using this website for the last 3-4 years. It was tremendous human coded tool

agents improve the task

I should try this out 🤯

i’d read the hacker’s notes first