A working philosopher’s manual for people who build memory systems, tools, and multi-agent worlds
Why a dead Austrian still has the floor
There is something quietly absurd about reaching for Philosophical Investigations when your stack trace is on fire. And yet — every time I sit down to design an agent memory system, a tool protocol, or a promise graph between cooperating agents, I notice the same shape underneath the work: somebody is trying to say what a word means, and somebody else is trying to ground that meaning in use. That shape is Wittgenstein’s. He never met an LLM, but he spent forty years sharpening the exact set of distinctions we now blunt our heads against.
Engineering folklore tends to grab one quote — “the meaning of a word is its use” — and stop there. It is not enough. Wittgenstein gave us a small, stubborn family of ideas that, taken seriously, change how you build agents:
Language-games: bounded activities with rules that confer meaning.
Forms of life: the cultural, embodied, practice-rich background a language-game presupposes.
Rule-following: the slippery question of what makes one continuation of a pattern correct.
The private language argument: there is no meaning without public criteria of correctness.
Family resemblances: concepts held together by overlapping similarities, not shared essences.
Pulled together, these ideas describe almost exactly what an agent does, and where it breaks. The rest of this essay is an attempt to take each of them down from the philosophy shelf and bolt it onto the architecture.
I. The early Wittgenstein, and why he was wrong in a productive way
The young Wittgenstein wrote the Tractatus Logico-Philosophicus (1921) under the spell of Frege and Russell. Language, on that picture, mirrors the world. Sentences are pictures of facts. Meaning is correspondence. The world has a logical structure; language has a logical structure; they fit.
Anyone who has built a symbolic AI system from the GOFAI era will recognise the program. Build an ontology, axiomatise the world, do inference over propositions, watch meaning emerge. It is a beautiful idea. It also did not work — not at the scale of human language, and not at the scale of agent behaviour either. The reason it did not work is the reason Wittgenstein himself walked away from it.
He came back twenty years later with Philosophical Investigations and quietly demolished his earlier self. Language is not a calculus. Words are not labels stuck on objects. The unit of meaning is not the proposition but the practice. This is the move that matters for us, because every modern agent system — from RAG pipelines to tool-using agents to multi-agent negotiations — is operating in the late-Wittgenstein register, whether its designers know it or not.
A language is not a class of sentences that can be formed from a set of axioms... but an open-ended series of interlocking language-games constituting a form of life.
That is one of the cleanest summaries of the shift, and once you hear it you cannot un-hear it when you are debugging an agent that hallucinates or a tool call that does the syntactically correct thing in the semantically wrong context.
II. Language-games: every prompt is a game with rules
A language-game (Sprachspiel), in Wittgenstein’s hands, is not a metaphor for chitchat. It is a unit of analysis. It is the smallest practice in which utterances acquire definite meaning. The famous opening of the Investigations — the builder who shouts “slab!” and the apprentice who brings a slab — is the minimal example. Four words. A trade. A relationship. A purpose. That is a language-game.
Wittgenstein’s larger list in §23 is the part I keep coming back to: reporting an event, speculating, forming and testing a hypothesis, making up a story, play-acting, guessing riddles, joking, translating, asking, thanking. Each of these is its own game with its own rules. The same sentence — “It’s going to rain” — performs differently in a weather report, a suspicious threat, a Slavic peasant proverb, a poem, or a model evaluation prompt. Same shape, different game, different meaning.
Now look at what an LLM agent actually does. A prompt is a game-rule specification. Role, goal, format, constraints, examples — all of this is the engineer trying, often clumsily, to specify which game the model should play. When the prompt is clear, the model performs. When the prompt is vague, the model picks some game from its training distribution — usually the most statistically average game in the neighbourhood — and you get the response we politely call a hallucination. It is not really a hallucination. It is the model playing a different game than the one you thought you were starting.
This reframing is more than rhetorical. It tells you where to spend engineering effort:
Prompt design is game-rule design. Stop writing prompts as wishes; write them as rules of a game.
RAG and tool calls are how you anchor the game in something other than the model’s prior. Retrieval changes which game the model can credibly play.
Evaluation is rule-conformance checking. The question is not “is the output good?” but “did the model play the game we specified?”
Most agent failures are game-mismatch failures, not capability failures. The model knows how to play some game competently; it just isn’t the one you wanted.
Once you internalise this, a lot of agent debugging becomes: what game does this prompt actually instantiate, given the model’s training distribution and the retrieved context? That is a much sharper question than why did the model say that?
III. Forms of life: the unwritten background
A language-game does not float free. It sits inside a form of life (Lebensform) — the broader weave of customs, bodies, institutions, and shared practices that makes the game intelligible in the first place. The slab-and-pillar game of the builders works because there is construction, there is labour, there is a society in which someone gets paid to bring slabs.
Wittgenstein only used the phrase a handful of times in the Investigations, but it is doing enormous work. It is the answer to the question, why do these rules count as rules? They count because a community lives that way. The form of life is what is given — not derivable, not justifiable from outside, simply the ground we stand on when we mean anything at all.
The modern agent has no form of life. This is the deep version of the grounding problem. An LLM has read about pouring tea but has never spilled any; it has read every recipe but never burned a pot; it has read about grief without ever losing anyone. Its sentences are, as Donald Clark puts it, echoes rather than experiences. “If a lion could talk, we could not understand him,” Wittgenstein wrote, because the lion’s form of life is alien to ours. An LLM’s situation is the inverse and stranger: it speaks our language fluently while inhabiting no form of life at all.
For agent architects, this has three concrete consequences.
First, embodiment is not optional in the long run. Systems like SayCan, Gato, or robotics-grounded agents are not just engineering improvements — they are attempts to build the beginnings of a form of life. The agent that can actually pour the tea has the start of a grip on what pour means. You may not be building robots, but every time you connect an agent to a real tool, a real database, a real consequence, you are attaching it to a tiny scrap of a form of life. That is what tool use is, philosophically.
Second, multi-agent systems are how forms of life get bootstrapped without bodies. When agents have to coordinate, deceive, negotiate, and hold each other accountable — Diplomacy, Werewolf, promise-graph cooperation — words start to acquire weight from the social consequences of using them. “I will support your attack” becomes a meaningful utterance not because of its syntax but because saying it and then defecting has costs. This is one reason Wittgenstein’s framework lights up around multi-agent research: a community is the minimum viable form of life.
Third, your domain is a form of life and you are the ethnographer. When you build an agent for medicine, law, finance, or trading-card cataloguing, you are not building a general-purpose linguistic engine; you are building a participant in a specific form of life with its own rules of correctness. The work of specifying that form of life — what counts as a good answer, what counts as escalation, what counts as a refusal — is the actual product. The model is the substrate. The form of life is the design.
IV. Rule-following: the regress that eats your guardrails
Here is where it gets uncomfortable.
Wittgenstein observed something disturbing about rules: any finite sequence of behaviour can be made to fit infinitely many rules. You teach a child to add 2 — 2, 4, 6, 8, ... — and at some point the child writes 1000, 1004, 1008. The child is, on one interpretation, following the rule “add 2 up to 1000, then add 4”. You wanted plus; for all your training was worth, you might have taught quus (Kripke’s term for the function that agrees with plus until some threshold and then diverges).
What rescues us from this regress in human life is not a deeper rule. There is no deeper rule. What rescues us is practice — the trained, communal, embodied capacity to take the next step the way “we” take it, where the “we” is a form of life. Rule-following bottoms out in what we do, not in what we know.
This is not a curiosity. This is the central problem of AI alignment, stated forty years before AI alignment had a name.
Every agent system that ships is, in effect, an attempt to specify plus and pray it doesn’t learn quus. Constitution prompts, RLHF, guardrails, evals, red-teaming — these are all moves in the same game: trying to pin down rule-following well enough that the agent generalises the way we want, not the way that also fits the data. And the unsettling Wittgensteinian observation is that there is no escape from this through more rules. You cannot specify your way out of the regress. The only exit is training and shared practice — exactly what RLHF and preference learning do, even if their practitioners never read the Investigations.
A few practical handholds fall out of this:
No specification is ever complete. Stop expecting the prompt or the policy to enumerate every case. The work is to ensure the agent generalises like a competent member of the practice, which means evaluation against a broad sample of the practice, not coverage of edge cases by fiat.
Behaviour, not introspection, is the criterion. An agent that says it understands the rule but acts otherwise is not following the rule. This points to performance-based alignment: judge by what the agent does in context, not by what it claims about its goals.
Adversarial generalisation is the quus problem. Jailbreaks and prompt injections are precisely cases where a rule that looked stable on the training distribution breaks at a threshold. Defence-in-depth here is not a moral failing of the model; it is a mathematical inevitability of finite rule-specification.
The Springer paper Philosophical Investigations into AI Alignment (2024) makes this argument explicitly: alignment is best understood as aligned performance between agents, not as alignment of inner states. We do not align by sharing mental pictures; we align by acting in compatible ways within a shared practice. That is a Wittgensteinian alignment story, and I find it more honest than the “values inside the model” story we usually tell ourselves.
V. The private language argument: why hallucinations are a category, not a bug
The most beautiful single move in late Wittgenstein, for me, is the private language argument. Imagine you keep a diary and write “S” whenever a certain inner sensation occurs. Could you, in principle, build a language whose terms are anchored only by your private acts of attention? Wittgenstein argues no. If “S” just means whatever you now feel, then whatever seems right is right, and the distinction between correct and incorrect use collapses. Without public criteria of correctness, there is no rule to follow, hence no meaning.
A purely private language is incoherent. Meaning needs publicity.
Now look at an LLM hallucinating. It produces fluent, confident, internally coherent text about a paper that does not exist. Within the model’s own activations, there is nothing flagging this as wrong. The model has no external criterion against which “the paper exists” can be tested. It is, in a precise Wittgensteinian sense, attempting to use language without public criteria of correctness. The hallucination is not a glitch. It is what language does when its public anchors are stripped away.
Marco Graziano, in his work on the Language-Game Description Language (LGDL), draws exactly this conclusion: hallucinations are the predictable behaviour of a system trying to mean things without participating in shared practices. The remedy is not bigger models. The remedy is grounding the agent in language-games with public criteria: tool calls whose results the agent must reckon with, retrieved documents the agent must cite, structured games where every move has a verifiable consequence.
This is the deepest reason RAG works. Not because retrieval magically inserts truth into the model. Because retrieval inserts a public criterion — something against which the model’s claims can be measured and corrected by a community of speakers (you, your evals, your other agents). Same for tool use: a tool returns a result the agent did not invent, and the agent must integrate that result. Same for multi-agent verification: another agent’s contradiction is a public criterion.
Build agents that are forced to participate in something public. That is the engineering form of the private language argument.
VI. Family resemblances: against the tyranny of definitions
One last idea, smaller in the literature but lovely in practice.
Wittgenstein noticed that some concepts — game, famously, but also agent, intelligence, memory, understanding — do not have a shared essence. Look at board games, sports, card games, video games, children’s games. There is no single feature they all share. There are family resemblances: overlapping similarities, like the features of relatives in a family photo. Some are competitive, some cooperative; some involve skill, some luck; some are physical, some symbolic. The concept is held together by a web of overlapping uses, not by a definition.
This is genuinely good news for graph database designers and ontologists. The dream of a clean, total, hierarchical taxonomy is a Tractarian dream. The actual structure of human concepts is closer to a graph with weighted similarity edges than to a tree of classes. Bipartite reification, polymorphic relations, similarity edges, embedding spaces, family-resemblance clustering — these are all post-Tractarian moves. Modern knowledge graphs are, accidentally, more Wittgensteinian than the symbolic systems they replaced.
It also tells you to stop arguing about definitions of AGI, agent, memory, intelligence. There is no clean definition. There are uses. Specify the use, build for the use, evaluate the use. The metaphysical fight over the word is, in the precise sense Wittgenstein meant, language going on holiday — language idling, disconnected from the work it normally does.
VII. A working architecture, philosophically grounded
Pulling the threads together, here is what a Wittgenstein-honest agent architecture looks like, in skeletal form. Treat this as a sketch, not a prescription.
1. Define the language-game explicitly.
Not as a system prompt. As a specification: roles, allowed moves, success conditions, failure conditions, escalation paths. Domain experts — clinicians, lawyers, traders, librarians — should be able to read it and recognise their practice. This is the LGDL idea: the form of life as an artifact you can version-control.
2. Bind every move to a public criterion.
Every claim the agent makes should, where possible, be tied to something it cannot privately fabricate: a tool result, a retrieved document, another agent’s confirmation, a database query, a checkable computation. The architecture should make it harder for the agent to hallucinate than to be honest. This is the private-language argument compiled into a runtime.
3. Treat memory as participation, not storage.
An agent’s memory is not a pile of facts. It is its accumulated participation in language-games — what it has done, what was confirmed or contradicted, what worked, what was challenged. A graph-shaped memory with edges like CONFIRMED_BY, CONTRADICTED_BY, USED_IN, PROMISED_TO is much closer to Wittgensteinian practice than a vector store of paraphrased “facts.” (This is also why promise graphs, social memory, and event-sourced agent histories feel so right philosophically: they record participation, not propositions.)
4. Make rule-following observable.
Don’t ask the agent whether it is following the rules; check whether its trajectory through the language-game conforms to the rules. Build evals that sample the practice broadly, not the edge cases narrowly. Expect quus moments at adversarial thresholds and design for graceful degradation, not exhaustive specification.
5. Build forms of life by composing agents.
A single agent has no form of life. A community of agents that cooperate, contradict, hold each other to promises, and accumulate shared history begins to. This is the most interesting frontier: forms of life are emergent properties of well-designed multi-agent systems, not features you can prompt into a single model.
6. Family-resemblance your ontology.
Stop writing rigid taxonomies. Use similarity edges, polymorphic relations, soft type systems, hybrid graph-plus-vector storage. Concepts will cluster by family resemblance whether you encode them that way or not; you might as well design with the grain of how meaning actually works.
VIII. The thing that is left over
There is a residue, after all this, that the framework does not dissolve.
Wittgenstein, in the Tractatus, ended with the famous line that whereof one cannot speak, thereof one must be silent. The late Wittgenstein softened this — he came to see that showing what cannot be said is the work of poetry, music, religion, ordinary life. There is a category of meaning that lives in the form of life itself and cannot be lifted out of it into propositions.
For agents, this is the limit. An agent can participate in language-games. With enough embodiment and enough community, it might inhabit something like a form of life. But the ground from which all of this springs in human beings — the bodies, the mortality, the having grown up somewhere, the smell of a particular tea you drank in a particular kitchen with a particular three-legged cat — that is not a feature you can specify. It is what makes specification possible at all.
This is not a reason for despair. It is a reason for humility. The agents we build are participants in our form of life, not bearers of their own. They mean what they mean because we hold them to standards drawn from our practices. The work of specifying those standards well — clearly, communally, in a way that respects the games being played and the lives that hold those games together — is, I think, the actual job. The model is the substrate. The form of life is the design. The form of life is, always, ours to keep tending.
Wittgenstein closed the Investigations with a remark that I want to leave you with, because it is the right note for this whole project:
What we are supplying are really remarks on the natural history of human beings — not curiosities, however, but rather observations on facts which no one has doubted, and which have only gone unremarked because they are always before our eyes.
The agents are before our eyes. The games are before our eyes. The forms of life are before our eyes. The work is to look at them clearly, and to build accordingly.
Further reading: Wittgenstein, Philosophical Investigations (esp. §§1–43, §§185–242, §§243–315); Saul Kripke, Wittgenstein on Rules and Private Language (1982); the recent Springer paper “Philosophical Investigations into AI Alignment: A Wittgensteinian Framework” (2024); Marco Graziano’s work on LGDL; and the MaKTO paper (arXiv 2501.14225) for an empirical Wittgensteinian study in Werewolf-playing multi-agent LLMs.
Originally published on Sovereign Agentic AI on Substack.