A repository-first working environment across every discipline, an onboarding path built for people who are new to it, and a written standard for telling the difference between what we think and what we have checked.
"Say which kind of claim it is."
AstroSpider is repository-first across every discipline - engineering, marketing, operations, research, hiring. Decisions, standards and plans live in version control where they can be reviewed, traced and reused. Prior Git or software experience is not required; new people get graduated permissions rather than a reading list, so the operations that are hard to undo stay out of reach until they are wanted. And every claim we write down carries a label saying how we know it.
Documents, decisions, standards and plans live in version control - not only for engineering, but for marketing, operations, research and hiring too. Scattered files and inboxes lose the reasoning; a repository keeps it somewhere it can be reviewed, traced and reused.
That is not a preference for developer tooling. It is the same thing we build for clients, applied to ourselves first: knowledge an AI system can actually retrieve and reason over, because it was captured deliberately rather than left wherever it landed. We are willing to make claims about AI readiness because we are living in the result of it daily.
The consequence is that bringing a person up to speed and bringing a machine up to speed stop being separate problems. Written conventions, decision records that carry an expiry condition, and a handover document per project are not bureaucracy. They are the mechanism by which somebody inherits context they were never present for.
Onboarding a person and priming an AI system turn out to be the same question: what is worth carrying forward, and who decides.
Prior Git or software-development experience is not required, including for roles where those skills would traditionally be uncommon. We teach the process. The goal is for people to get comfortable with versioned knowledge, documented decisions, review workflows and AI-assisted development - and we would rather work with someone who wants to learn that than someone who already knows every tool.
New people get a working environment with graduated permissions. The operations that are hard to reverse are not available yet; the ones that are safe to explore are. That is a property of the environment rather than a comment on the person, and it moves as they do. Being told to be careful is not a safety mechanism.
Instructions can be delivered one action at a time - each with the command, what it changes, how to check it worked, and how to undo it. That is a setting somebody chooses, not an assessment we make of them. It can be switched off mid-task, and it stays off.
Asking what a word means carries no cost. If our glossary did not cover it, the glossary has a bug. That rule matters more in a small company than a large one, because there are fewer people to ask and a higher price on quietly guessing.
We support neurodiversity and genuinely value people who approach problems differently. Conventional thinking is not assumed to be better simply because it is conventional. Different mental models, unconventional approaches and alternative ways of organizing information are where a lot of real innovation comes from. The bar stays the same for everyone: ideas should create value, solve problems, improve systems, or help people grow.
None of the above is a statement about who works here. It is a statement about what the environment is designed to accommodate - people arrive with different amounts of context and different preferences for how instructions reach them, and designing for that up front is cheaper than expecting everyone to adapt to one default.
Ideas get challenged; people do not get attacked. Leadership is not exempt from that. A team that cannot say "I think this is wrong" will ship the wrong thing on schedule.
Every assertion in our research notes carries a tag saying how we know it. Four gradations, applied per sentence rather than per document, because a document is almost never uniform - one paragraph of measured result sits happily next to three of speculation, and the reader deserves to be told which is which.
An untagged claim defaults to [belief], not to true. That default is the entire system. It makes assertion - the cheap move - the one that has to be justified, and it removes the comfortable option of saying something confidently and letting the reader supply the certainty.
Most of what we have written so far is [cited] and [belief]. That is what an honest early stage looks like: other people's results, credited to them, alongside our own ideas that have not yet earned the right to be stated flatly. A belief is not a weak claim - it is a candidate, and usually the most interesting thing on the page. The dull badge only means it has not been paid for yet.
"This doesn't work" is nearly useless. "This doesn't work on that model above roughly forty thousand tokens, and here is the output" is the whole value. So when a failure is written down, its boundary conditions are not optional - without them you get over-pruning, where a promising branch is abandoned because something specific to one environment was recorded as universal.
Every decision record ends with a Revisit if line: the circumstance under which it should be reopened. A decision with no stated expiry gets treated as permanent long after its premises have stopped holding, and nobody notices, because the reasoning that would have flagged it left with the person who made it.
What would count as failure is written down first. Deciding afterwards is how a null result quietly becomes "a promising direction," and it is very hard to catch yourself doing it in the moment.
If two people are using "memory" to mean two different things, the argument is fake and can run for hours. We fix the glossary first, then have the argument.
Nobody here needs to be right. Changing your mind mid-thread is a signal that the process is working, not a loss, and it gets said in the thread rather than fixed by a quiet edit.
The first thing this caught was our own copy. [observed]
Two work sessions drafting this page described things we hold as rules as though they were things we had already done. Both were caught by checking the claims against the repository rather than against memory. The tags did not do the checking - but they are what made the error nameable in one phrase, which is most of what made it fixable.
"AI has a memory problem" is too big to work on. Underneath it there are at least six separate problems, and they are neither equally hard nor equally crowded. Knowing which one you are standing in is most of the work.
Attention cost rises quadratically with length, so the window is capped rather than merely expensive. Extremely crowded - every lab is on it, with hardware nobody else has.
Of everything in a long working history, which parts are load-bearing later? Today a human decides, per session, with no articulable rule set. Thinly worked.
Stores are overwhelmingly append-only. A decision from March that was reversed in May sits there at equal weight and gets retrieved with equal confidence. Thinly worked.
Curated context is trapped - in one vendor's memory feature, one chat's scrollback, or one person's copy-paste habit. Moving it means rewriting it. Early, and heating up.
Approaches that failed are almost never captured, so they get suggested again. Humans have the same failure mode; we call it institutional memory loss. Nearly empty.
In a store, a checked result and something confidently invented in turn forty look identical. Both are text with an embedding. Moderate for retrieval, thin for agent memory.
That ranking is a judgment call rather than a result [belief], and it is the kind of thing we would rather be argued out of than quietly right about.
The two we keep returning to are salience and negative knowledge. Salience because it sits upstream of everything else - retrieval quality, graph quality and summary quality are all bounded by whether the right thing entered the store to begin with. Negative knowledge because it looks like the emptiest corner of the six, and because a dead end is unusually dense: "approach X fails because of Y" prunes an entire branch of the search space in one sentence, which positive facts rarely manage per token.
The counterintuitive part, and the reason more capacity does not close this: a very large window full of undifferentiated material can perform worse than a small one of curated context [belief]. If that holds, cheaper capacity has been delaying work on salience by making the symptom less painful.
The strongest argument against all of this, put as well as its proponents would put it: memory is retrieval, and retrieval scales. Every generation of longer context and cheaper embeddings has eaten a chunk of what looked like a curation problem. "Be smart about what we keep" has repeatedly lost to "keep it all and search better." Betting against that pattern means believing this time is structurally different, and the burden of proof sits with us.
Two things may survive it - search quality degrading with store size in a way capacity does not fix, and negative knowledge never entering the store at all however large it gets. Both need testing rather than asserting. [belief]
We have not published research. The questions above are ones we find interesting and are actively reading about - not results we have produced. Where we cite someone else's work, the credit is theirs and we say so.
If that reads as a strange thing to put on a company website, it is the same discipline as everything above it. The usefulness of a claim depends almost entirely on being able to tell what kind of claim it is.
If you are weighing up whether your organization's knowledge is in a state an AI system could safely use, that is a conversation worth having and it costs you nothing. If you are here because the way we work sounds like somewhere you would want to work, say so directly - that is the same inbox.
Get in Touch