We released Rúnir. https://t.co/vg6nqBovat
Rúnir is a discrete memory service for AI agents. Agents call it via HTTP once per turn: before the model replies, the service recalls the essential bits; afterward, it archives what deserves retention; finally, it tidies up. The rule is simple: just in time, just enough. That continuity isn’t a dump of your entire interaction history; it’s that narrow, curated fragment of retained knowledge worth placing into the prompt.
Memory here, then, is understood as living sense, not log scrapbooking. You save lived instances as semiotes (what was said and what was resolved), which feed semiosis (continuous interpretation of changes of context), which in turn modulate hexis (appropriate dispositional and affective stance in retrieval), while more permanent patterns coalesce into noema (stable sense-particles that survive the life of a conversation), and syndesis stitches the chain across time so future agents can still pick up the story. All writes are arbitrated (create/merge/supersede/skip) to allow facts to change without losing lineage, with “soft-inactivate” the path of least resistance and “hard-delete” an explicit choice. Retrieval is a hybrid mix (vector search, full text, recency, entity lookup), after which weakly-matched items are filtered to protect the token budget.
What we want to be distinctive here is separating the concerns: intelligence lives within Rúnir’s service layer, while Claude, Codex, Pi, Grok Build, OpenClaw, etc. act as thin clients. No mandated framework, no “session-starts-as-novel-as-memory” fiction, no expectation the model fabricates continuity from thin air when the service returns nothing. Ambient hooks support the ordinary “remember that I like …” style capture and recall; when you do want to know, explicit tools and deep surfaces (e.g., /search with citations) give straightforward answers even when the memory doesn’t know. Rúnir is for agent builders looking for continuity that compounds: less copy-pasting, fewer repeat explanations, more retained sense.
@manavvnotop Graphs are the current meta. Whether you're a fan of them or not, you're not doing it right.
You likely have some outdated notions of how things work. Take a couple weeks to learn what's happening currently.
I have plenty of tokens left at the end of the week, but I'm generally only working on several projects. I'm not running some sort of software factory or anything like that, just doing real work. I have a non-traditional workflow, so it's lasting all week for me. Currently, I'm using:
Opus 5.5 as the orchestrator
Luna 6 as the scout
Sol 6 as the builder
Grok 4.7 as the reviewer
Sadly, for Astra to be in a token burn state that can last all week, you have to run it with low effort. Anything higher than that and Astra sucks your usage down too fast. The other side of that is that, for a lot of things I'm working on, Astra doesn't function at the level I need it to when it's on low.
Right now, Astra isn't usable for me in a traditional “agentic dev” scenario unless I use it for one-offs, such as a UI design or a very complex novel problem I have, and in those cases, I’ll have it go back and forth with Opus 5.5 or Fable 5.1 to suss out each other’s weaknesses before building with the aforementioned workflow.
I made Jev do a stupider thing, be a NAND gate, so it can run a 4-bit ALU.
Adding 7 + 5 = 12 takes 116 gates, 7.6 seconds, $0.0018, and a compound 84% probability.
Billions of these will play Doom one day.
I've asked Opus 5.5 to make a video about it, Nolan style.
I think it's a mistake to nix reviewer agents because, at least with the current state of models, they will likely find at least one or two high-priority items if the task is long or complicated. And it's not always in the form of code that doesn't perform a certain task, but it's also in the area of drift.
Models aren't quite there yet. I'd suggest you keep the reviewer in, but your call.
I hard limit mine reviewers to two iterations as long as there are no P1s, if there are P2 and P3s left, I'll file tasks for those separately to fix later.
@crutchcorn Maybe it's time agents started writing code for disk efficiency, token efficiency, and for agent reads rather than human reads. Are you always reading your code, because most people don't anymore.
@theo I have too many deterministic evals and verifications in my workflows/workgraphs for my models to go off point this radically. I would suggest you do the same, regardless of which models you're using.
@forgebitz Use Grok build with CLI proxy, then you can use whatever you want. Just have your agent add your models to the config file in Grok Build, and you're good to go.
Understood. I suppose there are always going to be some friction when you've got people who don't really have a programming background. They vibe code everything, might have a YouTube channel or a following, and get to these places without having any real engineering foundation. If someone tries to pierce their influencer veil with legitimate concerns, they get offended, just like this guy did (or so it seemed). I'm still going to keep giving people recommendations and suggestions, especially where best practices are concerned.
To calibrate you all on which reasoning effort to use for Astra, know that GPT-6 Astra on low performs better than GPT-5.6 Sol on high.
If you were using high reasoning efforts with Sol and were happy, I suggest you move down to low or medium for Astra.
So far, what we're seeing is that Astra finishes most tasks in our typical workflows, which have a lot of deterministic tooling between each step of the workflow, are finishing in about half the time versus Fable. Token cost is also a lot lower - don't have a large enough sample size yet to give you an accurate reading, but I'd say it's at least 20-40% cheaper on token spend - some tasks, even larger gap.