Jim Martin

@theo 🥔 tomorrow?

It’s been over a month since a big new model dropped. Damn.

I need to find out how I can let someone know about this man, he’s always so friendly. You have a good one USPS, make sure you keep him 🙏

It’s the little things that make a big difference and can make a person smile. The mailman left my LMNT box at the door and on top of it was a Milk Bone 🦴 Every time my dog sees the mailman he goes crazy barking and this mailman was so nice to leave that for my dog. @USPS

@astro_reid Welcome home buddy. BTW, you and my bro, K-Mart, ran with the same group of guys in flight school. So proud of you guys! 🇺🇸

Mission complete ❤️❤️❤️

@amasad Nice! My company will for sure take a look. Our Databricks account reps are visiting us in Dallas next week to talk about AI/BI Does this send the Databricks data to the LLM? We deal with PHI and can't have our data being sent outside of Databricks

Replit now deploys directly to Databricks. Your apps run inside your Databricks environment while inheriting its security, governance, and data access. Beta is live. Enterprises are already building with it and seeing massive acceleration in BI and internal tools.

@steipete openclaw 2026.4.5 shipped with a busted package layout for Discord bits. 2026.4.5 tries to import buape/carbon from top-level dist files

@steipete God please let this work 🙏 Was just using Opus 4.6 as my orchestrator before canceling my subscription today. Hopefully 5.4 will now stop vomiting the same explanation 4 different ways and then a summary at the end and then telling me to do the work 😁

Anyone who wanna help me test the GPT 5.4 🦞upgrades: openclaw update --channel dev

@ismaguimarais @steipete @orangie Same, I experience this all the time with 5.4. Exactly like you posted here. I have to tell the damn thing it can fix it then it tells me it doesn’t have access to said thing when it clearly does.

GPT-5.4 kept explaining the problem instead of solving it. Here's what happened: I asked it to install the Obsidian CLI on my laptop through OpenClaw. It told me how to do it. I said just do it. It asked if I wanted it to. I said yes. Then it said "but your vault is on your laptop, not the VPS." Bro, you have SSH access through Tailscale. You know this. It tried winget, it failed, and it just... stopped. Switched to Opus. It SSH'd in, winget failed, Go wasn't installed, so it pulled the binary straight from GitHub, extracted it, dropped it in PATH, configured the vault, and verified it worked. Every time it hit a wall it just tried the next thing. No explaining, no asking, no stopping. That's the difference. One talks about blockers. The other burns through them.

what models are you using now for your openclaw since the Claude ban

Write code to help yourself, publish code to help others.

@digitalix So should we go with multiple DGX sparks or multiple MacStudios?

DGX Spark playbooks will soon include 2, 3, and 4 cluster node setups.

@digitalix @Dell Yes please 👋

Well, what do we have here? @Dell added this little beauty to their site.

@steipete @bradmillscan @vincent_koc Been working on this over the past month, making really good progress, but the long tail is where it gets rough.

@bradmillscan @vincent_koc Before making this default we gotta experiment and see how well this works, but yah, better memory is high up the wish list!

OpenClaw memory unlock! Forcefeed memories before responses! I just noticed that a new runtime hook was added to OpenClaw this week by @vincent_koc & others that will solve a lot of OpenClaw drift ... if someone builds a plugin for it! Who's building a communication plugin for OpenClaw using the new before_prompt_build hook? Using prependSystemContext and appendSystemContext, this allows you to inject extra instructions/directions BEFORE the agent builds the response. Here's how OpenClaw weights things: 1 Core system prompt 2 Plugins - prependSystemContext (NEW) -appendSystemContext 3 Agents.md - your custom rules 5 Tools > Skills - tool APIs and documentation 4 Workspace files - memory, playbooks/SOPs etc 5 long-term memory - retrieval from Obsidian/DB 6 session transcript - current convo context 7 your message - your request or you can think about it this way: prependSystemContext (NEW plugin hook) system prompt (you can't change this) appendSystemContext (NEW plugin hook) agents.md, bootstrap files tools & skills workspace files (memory, Playbooks & SOPs) conversation context prependContext (previous plugin hook) current message / request -- The biggest unlock of this addition to OpenClaw (which should be default behavior by OpenClaw cc @steipete) is a template communication plugin that turns on when a user activates memory_search & memory_get. The Default Comms Protocol Plugin should require the agent to use memory_search before asking the user a question. Majority of users who turn on openclaw's advanced semantic memory don't realize their bot is not using it. Even if you add a hard rule in agents.md that the agent must use it's memory tools before asking questions, majority of the time it does not use the tool. This is surfaced when you ask your bot to read the logs and show how many times it uses memory_search and memory_get over the last 24 hours. usually the answer is close to zero. Anyway, with this new runtime hook exposed, you can now really tune the kinks out of how your openclaw agent communicates with you. Don't want your agent to offer you "if you like I can ..." rabbit holes? Don't want your agent to say "good catch" ? Don't want your agent to ask you things it knows already? Don't want your agent to say "you're right to call that out..." Train it out with a comms plugin that has prompt weight above everything else, and then block messages that come back to you violating the comms protocol & force the agent to rewrite them to spec using message_sending hook for outbound filtering. This can also be used more practically for other things like token caching, model routing & multi-agent routing. You can now route messages to models more effectively to save money, and catch context-switching messages that are sent to the wrong agent before it bloats the context window of the main agent... Is anyone building on this?

@OpenAI The latest Codex app release is busted. It acts like it's frozen, and as it's working, nothing is output to the screen. You have to click around to get the screen to refresh.

My OpenClaw Agent is smarter than yours 😉 github.com/agenr-ai/agenr @openclaw

Children are our greatest natural resource. --Herbert Hoover

Now I just need some devs to help me keep it going. github.com/agenr-ai/agenr

Just watched my OpenClaw AI assistant recall a specific bug from a dev session that was deleted weeks ago. No notes, no docs, just the knowledge in her human like memory this is why i built AGENR github.com/agenr-ai/agenr. Memory that outlives the context window.

Built it for myself. Used it while building it. Tonight it recalled a design decision from a session that was deleted weeks ago, knowledge that survived the context reset. The thing remembers how it was built. 🤯

It's not Mem0. It's not RAG. The design pulls from cognitive science. Forgetting curves, spaced repetition, elaborative encoding applied to AI agent memory. Structured knowledge types. Importance scoring. multi-tier dedup. Entries that decay, consolidate, and surface smarter over time. The data lives in your SQLite db. The intelligence lives in the design.