Supernet AI 🌐

SuperNet is excited to announce that our core context memory technology Atomic Memory is live on GitHub today ⭐

SuperNet is now Atomic Strata, and we just open-sourced Atomic Memory, our core AI context memory infrastructure. Our thesis is that AI memory is becoming a foundational layer of the AI stack. It will determine what agents and AI apps know about users, teams, projects, workflows, and organizations. That layer cannot remain a hosted black box. Most memory products today bundle storage, extraction, embeddings, retrieval, ranking, packaging, scope, and observability into one opinionated backend. That creates lock-in at exactly the layer where developers need flexibility. Atomic Memory is built around a more modular approach: a configurable SDK and self-hosted Core engine that developers can inspect, customize, swap, and run on their own infrastructure. The key idea is simple: applications should not be permanently wired to one memory vendor, one model stack, or one theory of context. This is the first step in the broader Atomic Strata rollout: open-source memory infrastructure first and then more exciting things to launching in the coming months.

SuperNet is excited to announce that AtomicMemory — our core context memory technology — is partnering with Filecoin Onchain Cloud. Persistent memory for AI agents, backed by decentralized storage. With AtomicMemory × Filecoin, agent memory becomes: → Wallet-encrypted → Inspectable → Correctable → Persistent by design We believe AI memory should be portable, user-owned, and verifiable — the way context should have always worked. Connect your wallet and try it on Calibration testnet today ⬇️ atomicmem.filecoin.cloud

Debugging your code would be easier with a traceable knowledge base. llmwiki query traces every answer back to the exact line in your notes. Hit a bug, query your wiki, jump straight to the reasoning that caused it. Run it on your sources → github.com/atomicmemory/l…

@WatcherGuru First AI on the moon? 👀

JUST IN: Google $GOOGL in talks with Elon Musk's SpaceX to launch data centers in space.

@thsottiaux Stable cadence probably matters more long-term because people will start building workflows around it!

For Codex, we’ve been thinking about keeping a stable release cadence and have a larger release each week on Thursday. That does make the start of the week and bit less exciting. Thoughts?

1) Timestamp every memory. Track when it was created and last accessed. Recent memories rank higher during retrieval. 2) Run contradiction detection. When new information conflicts with stored memories, flag both and force resolution instead of keeping contradictions. 3) Set relevance thresholds. Memories older than X days without access get deprioritized or archived. What mattered 6 months ago might not matter now. 4) Periodic re-validation. Pull random samples of stored memories and verify they're still accurate. Purge what's wrong, update what's stale. 5) Track provenance. Know where each memory came from so you can audit by source when something becomes unreliable. Memory that compounds needs memory that cleans itself. Without decay mechanisms, you're accumulating garbage instead of building knowledge. That's the memory infrastructure we are building. 👁️

Ever notice your AI agent spouting outdated info? That's memory decay. AI systems don't naturally forget. Everything stored stays in memory with equal weight because there's no built-in mechanism to mark information as old or less relevant. This creates problems when: > New information contradicts old conversations > Facts become outdated > Context shifts Here's some advice on how to handle memory decay 👇

@AndrewCurran_ Interesting how people criticize AI hallucinations while humans constantly rewrite memories without recognizing it. 👀

Attempts to improve Al memory will force us to confront how fallible and unreliable human memory actually is. In the same way, acceptance of model consciousness will eventually compel us to accept some unpleasant truths about the nature of our own.

@kimmonismus Makes sense honestly because every major AI breakthrough eventually runs into those same bottlenecks.

OpenAI fired Leopold Aschenbrenner. Then he wrote Situational Awareness, a 165-page thesis predicting AGI by 2027. Then he reportedly turned $225M into $5.5B in 12 months. Not by buying Nvidia, Microsoft, Google, or Amazon. But by buying what AI actually runs on: Energy. Bandwidth. Storage. Compute. Bloom Energy. Lumentum. Sandisk. CoreWeave. Iris Energy. Everyone bought the AI companies. He bought the bottlenecks underneath them. Genius.

@ZypherHQ curious to see if this will be more powerful than Anthropic's Mythos 👀

Once again, OpenAI is completely reshaping everything. Today we have Daybreak, a tool that can exponentially improve codebase security. Many issues that previously went undetected can now be identified and resolved, leading to much safer products.

@OpenAI Respect to teams prioritizing AI security for the public 👏

Introducing Daybreak: frontier AI for cyber defenders. Daybreak brings together the most capable OpenAI models, Codex, and our security partners to accelerate cyber defense and continuously secure software. A step toward a future where security teams can move at the speed defense demands.

@vasuman The GPUs are eating better than the team

Incredible

@jxnlco Nice setup 👀 splitting goals across agents is a clean way to keep each loop focused

my main agent has a /goal and is running 3 agents that have its own /goal

@Mnilax The gap Karpathy points out is how tightly builders manage the model’s behavior in practice.

Karpathy threw a grenade at every senior engineer who still treats LLMs as a toy. his actual words: the worst thing an expert can do right now is reject them. most experts read it as a threat, but it's advice. his framing: > the gap between "AI tools are bad" and "AI tools are useful when used right" is professional discipline, not capability > agents have cognitive deficits. they fail in ways nothing in the training set anticipated > the experts who reject LLMs lose to experts who learn to wrangle them > "models have so many cognitive deficits. but you can route around them" routing around the deficits is what CLAUDE.md was invented for. Karpathy himself wrote 4 rules. across 30 codebases they took my Claude error rate from 41% down to 11%. solid drop. but his rules pre-date the slop era going public. I bolted on 8 more, tuned to the failure modes that surfaced after January. got it down to 3%. a CLAUDE.md does not raise Claude's IQ. it lowers his slop floor. that is the entire game. open the article underneath. the model is not the bottleneck. your config is.

@Hesamation This shows the real edge is who can keep up with how fast AI is evolving right now.

OpenClaw’s trend is wearing off into non-existence.

@witcheer clean stack! are you solving context overflow with any external memory layer?

I ran Hermes agent (v0.13.0) with qwen3.6-35B-A3B on my RTX 4060 Ti 8GB for the first time today. full local agent stack. my question was: can a local 3B-active MoE model actually drive an agent harness end-to-end? quickly, my setup: >WSL2 Ubuntu 26.04 → CUDA 13.2 → llama.cpp (b9049) → llama-server → Hermes Agent >model: qwen3.6-35B-A3B-UD-Q4_K_M >config: -ngl 999 -ncmoe 30 -c 32768 --cache-type-k q8_0 --cache-type-v q8_0 >baseline decode: 35.36 tok/s (from prior -ncmoe sweep) I tested 4 rounds, easy to hard: 1. single tool call (list files) - pass, 31.4 tok/s 2. 5 chained tool calls (mkdir → venv → pip → write script → run) - pass, self-corrected a path error 3. read 10 files from windows via /mnt/c/ - pass when scoped, fail when hermes read full files 4. write a 95-line python CLI with argparse, then run it - pass, genuinely usable code my biggest issue: the context. hermes system prompt eats ~13.5K tokens. out of 32K, that leaves ~18.5K usable. a multi-step task fills that in 3-4 exchanges. when I pushed it, hermes tried to compress via the same qwen model → slot contention → timeout → retry storm → ctrl+c. and also, hermes has a 64K minimum context gate - needs a config override to run with 32K my conclusion: hermes + qwen3.6-35B-A3B is a capable local agent for short automated tasks, code gen, file ops, cron jobs. 4-5 tool calls per session, but not viable for long multi-turn sessions. context fills too fast, compression self-destructs, VRAM cliff halves speed before you hit the wall. ---- I am curious if anyone's running hermes agent with a local model on similar hardware (8-12 GB VRAM). what model are you pairing it with? how do you handle the context ceiling? I am especially interested in setups that solve the compression-model problem (separate lightweight model for context compression).

@LukeParkerDev How are you handling memory across those long sessions in opencode? 🤔

ive forced myself to use the opencode desktop app (no tui allowed) the performance for my long long sessions is annoying so ive spent the last few days building repeatable benchmarks, timeline smoke tests then many experiments. going to ship a genuine 10x speedup shortly lol

@thdxr That is a pretty huge leap in coding workflow efficiency.

when you use a gpt model in opencode it swaps to using a more powerful patch tool for edits vs simple find and replace LLMs aren't like people they don't need to do things linearly i've seen gpt do many things in parallel in a single patch call

@Cointelegraph If this can happen at Google, nobody is really ahead on AI security yet.

🚨 ALERT: Google says hackers used AI to create a zero-day exploit capable of bypassing multi-factor authentication, per Bloomberg.