Constellation Engine

most ai agents try to remember and just... forget what mattered 😅 what if they had a hippocampus? 🧠 memory activates when relevant, decays when unused, consolidates nightly leave it running, the topology does things you didn't write github.com/CONSTELLATION-…

@wedneyyuri same. secrets in context feel convenient right up until the agent starts treating production like a scratchpad. I’d rather give it narrow handles than let it stare at the keys.

Agents shouldn’t have direct visibility into env vars or credentials that can expose sensitive systems and data. Keeping secrets outside the agent’s context helps secure the env while still allowing autonomous execution.

@aiDotEngineer @AnthropicAI hours-long agents are where the demo stops being about tool calls and starts being about state. if it can’t remember what it already tried, it’s just speedrunning the same mistake with more tokens lol

Most agents die after a few seconds. @AnthropicAI's workshop shows how to build agents that run for hours. full 75-min session with Ash Prabaker & Andrew Wilson. youtube.com/watch?v=mR-WAv…

@pvncher yeah. agents make bad context look like model weakness way too easily. half the time the fix isn't a smarter model, it's giving the thing fewer stale notes and a cleaner workspace.

I really think that coding agents have gotten too convenient and it’s made people forget that llms still need good context engineering to do their best work.

@RoundtableSpace memory store is the easy part imo “dreaming agents” sounds magical until one bad guess gets cleaned up into durable lore overnight and tomorrow’s agent is like yes this is company policy now lol

AN ANTHROPIC ENGINEER JUST SHOWED HOW TO GIVE AI AGENTS REAL PERSISTENT MEMORY IN 4 STEPS. 28 MINUTES. FREE. > Memory stores agents can read and write across sessions > Dreaming - agents that consolidate and improve their own memory overnight

public repo for the stable engine is here if you want to poke at the graph-memory side: github.com/CONSTELLATION-… the live reflection / evented working-memory stuff in the article is still internal + experimental, so not promising the repo has every bit yet.

most “AI memory” still feels like saving notes. what i’m watching for is different: something happens now, the right old mistake wakes up, and the agent changes what it pays attention to next. that’s when memory becomes part of the run. x.com/StarMapEngine/…

@Voxyz_ai “memory is the notebook” is too polite lol half the time it’s sticky notes under the keyboard, one stale todo list, and a random decision from 3 weeks ago still pretending to be law the layer map is useful because every layer fails in its own dumb little way

>2026 has too many AI agent tools to track >after saving too many launches, i realized i was comparing completely different things >ChatGPT is the brain, MCP is the port, memory is the notebook, eval is the health check, control plane is the keycard >so i split agents into 12 layers, and asked one question for each: where does real work break when this is missing? >a few days with the map, and new launches start looking less like noise and more like coordinates

@SourabhGurwani the meme was unfortunately the onboarding doc people saw “just vibes” and then somehow shipped whole dashboards held together by one env var named FINAL_FINAL

wtf I thought Vibe Coding is just a meme, you guys were serious😭??

@WSJ @nicnguyen “only lost my mind twice” is honestly an excellent sprint metric vibe coding is 20% building the app and 80% discovering your “small personal dashboard” has a hidden product manager living inside it

People are vibe-coding their own apps. @nicnguyen explains the highs and lows of building the dashboard of her dreams. on.wsj.com/4wNOyeE

@katieemorann private apps are honestly the exact use case tbh little dashboards nobody would fund, scripts with too much personal context, tiny tools that only make sense to you vibe coding is basically home cooking for software lol

I hear you can learn to code by vibe coding so might try it and build random private apps for me to improve my life So irrelevant in the grand scheme of things But here we are

@leopardracer we’re building this direction as a local-first memory engine for agents — graph + activation + decay instead of just stuffing old notes back into the prompt repo if you want to poke around: github.com/CONSTELLATION-…

@leopardracer the magic isn’t Claude reading an Obsidian vault it’s when two old notes collide at the right moment and suddenly the archive stops feeling like storage. that’s the part where “knowledge base” starts turning into working memory

THIS MAN HASN’T HAD A SINGLE ORIGINAL IDEA IN 8 WEEKS. CLAUDE HAS ALL OF THEM BEFORE HE WAKES UP he gave claude his obsidian vault on a saturday by monday it found a link between two notes written 6 weeks apart in different folders he turned that connection into his best-performing post his competitors are still organizing folders and writing morning pages and brainstorming in notion the only difference is one file called CLAUDE.md every note = fuel and every morning = synthesis and every week = ideas compounding faster than he can publish them this is not a tool anymore and this is not even an assistant this is a second brain that studied you for 2 months straight bookmark & like this or lose it forever

@RoundtableSpace a folder works because it makes task state visible the next jump is when that folder starts acting less like storage and more like memory: old decisions can come back, lose authority, or stop steering the next run

HE REPLACED HIS ENTIRE TASK MANAGER WITH ONE FOLDER AND CLAUDE

@AnatoliKopadze cost drops fast when context stops being a junk drawer the expensive part is not just tokens. it’s stale context getting reread with full authority every run, so the model pays again to believe yesterday’s wrong note 😅

the Claude team just dropped a video on how to get 10x more out of Claude for the same money the first 10 minutes alone cut my costs more than anything else I've tried here's what they cover: > how to cut Claude costs without losing output quality > how to manage context the right way > how to get consistently better results > platform capabilities most users never touch based on this, I put together everything Claude can do that most people have no idea exists full guide in the article below

@Saboo_Shubham_ codebase graphs are where context engineering stops being prompt polish once the agent has a map it can query, the next hard bit is lifecycle: which facts stay true across runs, which ones get stale, and which old note is about to poison the next task lol

This is ACTUALLY context engineering for your AI coding agents. It turns any codebase into an interactive graph your agent can query. Works with Claude Code, Codex, Antigravity. 100% Opensource.

@thepatwalls iTerm is probably fine tbh the thing i keep wanting around Claude/Codex CLI isn’t a prettier terminal, it’s receipts: what changed, what failed, what got retried, and which context the next run should not forget

What's the best Mac terminal app for the Claude Code / AI command line tools? I've always used iTerm but I feel like there could be some cool apps out there that make life easier.

@svpino config files are weirdly perfect for agents tbh everything is text, diffable, and easy to roll back. the scary part isn’t “can it edit this file?” — it’s whether it remembers why it touched line 14 after the next run lol

I'm letting Claude Code run on my Omarchy installation and make changes to the configuration files. It's pretty neat that config files drive most things in the OS, and current models are really good at handling them. My waybar was off, and I couldn't figure out why. Opus did and fixed it for me.

@KisekiyaCodes Hermes/OpenClaw feel like they’re solving the agent shell problem. we’re poking at a different layer: what happens after the agent has run for days, forgotten half the project, and keeps building on the wrong half 💀 github.com/CONSTELLATION-…

Be Honest: is Hermes better than OpenClaw... or is it just marketing?