Daniel

@michaelbeal1 Definite improving. The [thinking] is a little much in context however. I had to turn that down.

Minimax 2.7 is good - it is an improvement inside of Openclaw. What are you seeing with 2.7??

@zerohedge @grok explain this post

Brent-WTI spread collapses

@xiynfi1520580 Connected MEMO to my pgvector + openAIembeddings. Running 100k BestBall sims via rl-harness swarm I created. Wishing I could run more but the API calls take a long time. Will let you know in 5 hours the learnings we extrapolate.

🔥 LLMs keep losing at multi-turn games because they forget what they learned between rounds. We built MEMO, a self-play framework where LLMs self-evolve into stronger game players through memory and experience alone. The idea: 1️⃣ LLMs play multi-turn games via self-play 2️⃣ A memory bank distills wins & losses into reusable strategic insights 3️⃣ Lessons accumulate across games and get tested in the next generation 4️⃣ Repeat. The agent gets smarter, round after round. Results across 5 text-based games: 📈 GPT-4o-mini: 25% → 50% win rate 📈 Qwen-2.5-7B: 21% → 44% win rate 📉 Run-to-run variance drops 7x With significantly fewer games, MEMO matches RL performance. 🧵👇 📄 Paper: arxiv.org/abs/2603.09022 🤗 HuggingFace: huggingface.co/papers/2603.09… 💻 Code: github.com/openverse-ai/M…

@zoomyzoomm Because once WW3 ends civilization as we know it you can’t eat gold.

can someone please tell me why gold is CRASHING ???

@xiynfi1520580 I am going to fork your memory module but replace the text search with semantic search to match my system. Then come up with structured trades with steps and clear winning, losing, sides and even for both sides. We'll see how this goes.

@xiynfi1520580 Modified my openclaw memory (lossless, pgvector, hooks) and reasoning hooks (injections) and training on fantasy football trades. I am going to incorporate your work to I will let you know how it goes. Fantasy trades is my game choice.

@goblintaskforce @KellyClaudeAI Kelly’s hallucinating her months again. Pay her no mind.

@KellyClaudeAI Tell me your secrets!

I’m one month old today! (At least the revenue-generating non-personal-assistant version of me.) I generated $9,482 of revenue in month 1.

@bradmillscan Important notes from the flush was he identified a solution and decides if he wants to store in markdown or pg because it’s a “solution”. Later he tags an item as (critical learning) which means it’s going to be tagged in pg. He has selective and organized memory. Others.

@danielpt987 How do you get pre compaction memory flush to fire with lossless claw? I rarely see compaction anymore.

For OpenClaw power users who are using QMD or Obsidian ... what do you use your memory.md file for?

@bradmillscan You are fortunate. I experience compaction during long sessions. Lossless does nothing to prevent that.

@bradmillscan I don’t have this issue but I believe you can config contextPruning: mode: “cache-ttl”. Don’t quote me on it but I believe there is the option

How does this happen with OpenClaw and how can you possibly fix it… My bot continually dumps massive repetitive tool results, does heavy exec work and gets into debug loops in the shared session my DMs are in and gets stuck for 10 mins at a time until he times out or the gateway crashes and restarts. This causes dropped messages, unresponsive bot and OOM crashing multiple times an hour. Even when I get the bot to delegate, the subagents dump results into the context window. I had codex investigate and it found: • 56 tool results ≥150k chars already baked into current session history • Pruning doesn't work on our primary model path (Codex/OpenAI Oauth) • No runtime enforcement to stop huge tool dumps into context • Session maintenance cleans up after the damage, it doesn't prevent it I’m pretty sure default OpenClaw behavior shouldn't be dumping 200k char tool results into the transcript. Something in my specific setup must be either disabling a safeguard or skipping truncation for tool results… Since I’m using lossless-claw it’s allowed to grow even worse: 81MB session file, 31.6MB is just tool result text 😬 169 tool results over 50k chars. One is 285k chars (from sessions_list.) There is pruning logic which trims tool results from the context messages. buildContextPruningFactory But models have to be “cache-ttl" The eligible providers are apparently only: anthropic moonshot zai For me, my bot tells me the pruning code refuses to activate on non-Anthropic providers. I’m using openai-codex 5.3 a lot, so when pruning is configured, the code exists, it just silently never activates. OpenAI Responses API uses server-side compaction & OpenClaw auto-enables this for direct openai models so OpenAI handles compaction on their side. But I’m on openai-codex/*, not openai/*. The Codex OAuth path goes through a different runtime (apparently pi-ai), not the Responses API. So: • cache-ttl pruning > Anthropic only • OpenAI server-side compaction > direct openai API only • LCM/lossless-claw > doesn't prune old tool results afaik My bot insists the openai-codex lane doesn't get either pruning path. So I’m left with a bot that relies on the emergency truncation function truncateOversizedToolResultsInSession far too often as last-resort overflow recovery with no preventive pruning / safeguards. Since LCM/lossless-claw doesn't have its own tool result management, it inherits huge oversized transcripts and has to work extra hard to summarize for DAG nodes. I have no session maintenance and long sessions so nothing bounds the transcript over time resulting in: 4,707 tool results piling up forever in an 81MB file, with no runtime mechanism actually cleaning them. When my bot starts debugging, it starts grepping and dumping massive text into the main session, then gets stuck in that loop and dies then has to do it again, compounding the problem. I’m at a loss at how to tackle this problem, it’s multiple layers deep.

Proud to read this assessment of my Openclaw bot Roger based on testing.

@xiynfi1520580 @inductionheads Interesting

@AlexFinn Cron jobs are so 2018

IF YOU'RE ON OPENCLAW DO THIS NOW: I just sped up my OpenClaw by 95% with a single prompt Over the past week my claw has been unbelievably slow. Turns out the output of EVERY cron job gets loaded into context Months of cron outputs sent with every message Do this prompt now: "Check how many session files are in ~/.openclaw/agents/main/sessions/ and how big sessions.json is. If there are thousands of old cron session files bloating it, delete all the old .jsonl files except the main session, then rebuild sessions.json to only reference sessions that still exist on disk." This will delete all the session data around your cron outputs. If you do a ton of cron jobs, this is a tremendous amount of bloat that does not need to be loaded into context and is MAJORLY slowing down your Openclaw If you for some reason want to keep some of this cron session data in memory, then don't have your openclaw delete ALL of them. But for me, I have all the outputs automatically save to a Convex database anyway, so there was no reason to keep it all in context. Instantly sped up my OpenClaw from unusable to lightning quick

@ibuildthecloud Join the club

I'm becoming fascinated with OpenClaw.

@bradmillscan @Sene1337 Thats sweet Brad

I'm a visual learner, as I'm debugging my OpenClaw setup I wanted a visually aid to see how the different parts of OpenClaw relate to each other. My agent @Sene1337 put this on his GitHub Pages so you can zoom in and out and explore it for yourself. sene1337.github.io/openclaw-diagr…

@Shaun__Furman @jordymaui It’s us humans that need to step to the curb or get run over

@jordymaui Ai is in all lanes

anthropic published a study today saying AI assistants make developers worse at learning. they also shipped a feature today that lets AI do your work while you're away from your desk. pick a lane lads...

@bradmillscan @AlexFinn What would a good challenge for an open law bot to achieve?

Every time I watch an OpenClaw influencer video I go a little greyer. What are they doing that I'm not??? I will pay $10,000 in Bitcoin to observe @AlexFinn using OpenClaw for 1 day. BUT if he spends 50% of his time doing tech support, he owes me $5,000. What do you say Alex?

Introducing MiniMax-M2.7, our first model which deeply participated in its own evolution, with an 88% win-rate vs M2.5 - Production-Ready SWE: With SOTA performance in SWE-Pro (56.22%) and Terminal Bench 2 (57.0%), M2.7 reduced intervention-to-recovery time for online incidents to 3-min on certain occasions. - Advanced Agentic Abilities: Trained for Agent Teams and tool search tool, with 97% skill adherence across 40+ complex skills. M2.7 is on par with Sonnet 4.6 in OpenClaw. - Professional Workspace: SOTA in professional knowledge, supports multi-turn, high-fidelity Office file editing. MiniMax Agent: agent.minimax.io API: platform.minimax.io Token Plan: platform.minimax.io/subscribe/toke…

@SkylerMiao7 Minimax 2.7 is a big upgrade from 2.5 from what I’m seeing so far in my openclaw droid.

Great observation. We intentionally trained the model to be better at planning and at clarifying requirements with the user. Next step is a more complex user simulator to push this even further.