This is the argument page for Operation Scarecrow, ProductHarness's section on AI memory and Company Brains, picking up right after the section's opening piece, If Only We Had a Brain — the piece everything else here assumes a reader has seen. It restates and extends the case made in The Backplane Isn't One Memory. It's Four., adding one thing that essay didn't have room for: the fact that this four-way split has been discovered twice, independently, by two fields that weren't talking to each other. Source citations at the foot of this piece.
Ask a vendor selling a Company Brain product what it stores, and the honest answer is usually one thing: an index. Slack messages, wiki pages, tickets, documents, all embedded and made searchable, so an agent can retrieve whatever seems most relevant to the question in front of it. That's a real capability, and it's not nothing. It's also, almost without exception, the entire architecture — one store, one retrieval mechanism, one governance rule, applied uniformly to everything the organization knows.
That uniformity is the actual reason most of these initiatives stall after an impressive-looking pilot. Not because the retrieval is bad, but because "everything the organization knows" isn't one kind of thing, and pretending it is works fine right up until it doesn't.
What memory actually is, once you look closely
Cognitive science settled this decades ago, for individual minds rather than organizations. Working memory is the transient scratchpad — what you're holding onto right now, for the task in front of you, gone in minutes without reinforcement. Long-term memory splits again, and the split matters: semantic memory holds general facts stripped of the moment you learned them (what a word means, what category something belongs to), episodic memory holds specific, dated events (the meeting where the decision got made, not just the decision), and procedural memory holds skills — how to do something, encoded as behavior you can execute without being able to fully narrate it.
These aren't shelves for the same kind of content sorted by topic. They're retrieved differently, updated differently, and built differently. A semantic fact doesn't usually arrive fully formed — it's abstracted out of episodic memory through consolidation, a real process in which one specific experience, retrieved and reinforced enough times across separate contexts, generalizes into a fact that no longer needs the original episode attached to it. Complementary Learning Systems theory names the mechanism directly: episodic memory is a fast-learning system tied to the hippocampus, capturing single experiences; repeated episodes consolidate over time into the slow-learning, generalized knowledge held in the neocortex. Consolidation is triggered by repetition across separate instances, not by one person deciding something feels important. That distinction — earned through independent reinforcement, not asserted by a single confident judgment — is the hinge the rest of this section swings on.
Organizations already had a name for this
None of the above is a new metaphor applied to business for the first time. Jeffrey Walsh and Gerardo Ungson gave organizational memory its first integrative framework in 1991, in the Academy of Management Review, and it's been cited for over three decades since. They described five "retention bins" where organizations store what they know: individual memory, culture, transformations, structures, and ecology. Two of those — culture and ecology — are retrieved by humans tolerating ambiguity well enough to read a room or notice what's sitting on a shared desk, and have no clean equivalent in anything an AI agent can query. The other three map, imperfectly but usefully, onto the same four-way cognitive split: a specific decision with its rationale is episodic; a settled definition or hard constraint is semantic; a codified process is procedural; an in-flight, unresolved status is working memory. That mapping — organizational content read through individual cognitive categories — is the whole argument of the companion essay this page extends, and it's not repeated in full here.
What's worth adding is a second, independent line of evidence that arrived at nearly the same place from an entirely different direction, decades later.
Two lineages, one split
In 2023, a Princeton- and DeepMind-affiliated team — Theodore Sumers, Shunyu Yao, Karthik Narasimhan, and Thomas Griffiths — published "Cognitive Architectures for Language Agents," known in the field as CoALA. It decomposes a single AI agent's cognition into working, episodic, semantic, and procedural memory, explicitly inherited from cognitive science's memory taxonomy — the same source Walsh and Ungson drew on, applied one layer earlier than either of them applied it. Where Walsh and Ungson mapped cognitive categories onto an organization's retention mechanisms, CoALA mapped the identical four categories onto a single agent's internal architecture: working memory as transient reasoning variables held during one episode of thought, episodic memory as a timestamped log of past experience, semantic memory as generalized declarative facts, procedural memory as executable routines and learned skills.
Nobody on the CoALA team had a reason to have read a 1991 management-theory paper, and nobody applying Walsh and Ungson to a modern organization was thinking about agent cognitive architecture. Two fields, thirty-two years apart, with no citation path between them, independently converged on the same four-way partition of what "memory" means for a reasoning system — one studying organizations, the other studying agents. That's worth taking seriously as something stronger than a nice analogy. When one field discovers a distinction, it might be an artifact of that field's own vocabulary. When two unrelated fields discover the same distinction from opposite directions, the more honest conclusion is that the distinction is tracking something real, not something either field invented to make its own argument tidier.
This page's application — reading an organization's memory through the lens of what its AI agents specifically need to consume — is a third pass at the same split, one level removed from both originals. It inherits Walsh and Ungson's organizational scope and CoALA's agent-facing framing, which is exactly why the honest caveat below matters: a framework validated twice is still a framework being extended a third time into territory neither original paper was built for.
Thirty-two years apart, no citation path between them, one studying organizations and the other studying agents — and they landed on the same four-way partition.
Why "one index" breaks, specifically
Put the two lineages together and the practical claim sharpens considerably. Each of the four memory types has a different right answer to four separate questions an organization's AI agents actually need answered: how cheap should this be to write, how should a change to it get reviewed before it's trusted, how long should it live before it's pruned or promoted, and how confidently should an agent state something it retrieved from this tier.
Working memory should be nearly free to write and just as free to discard — a decision queue or in-flight status that dies with the workflow that produced it, no review gate, because governance ceremony on something meant to be disposable is the wrong amount of ceremony applied to the wrong object. Episodic memory is a precedent, not a law: worth recording with a timestamp and a reason, genuinely useful to the next person who hits the same question, but not yet something an agent should state as settled fact. Semantic memory is the opposite end of the spectrum — the glossary entry, the entity definition, the hard constraint — expensive to get wrong precisely because it has the widest blast radius in the system, which is why it's the one tier that actually needs a named steward with real approval authority rather than a good intention. Procedural memory sits apart from the other three entirely: getting it wrong doesn't look like a misunderstanding, it looks like a bug, which is why it belongs under the same engineering review a codebase already gets rather than under content or product review.
An architecture built as one undifferentiated index answers all four of those questions the same way for everything it stores, because it has no mechanism for asking which tier a given piece of content actually belongs to. That's fine for the first weeks of a pilot, when most of what an agent needs is genuinely working-memory-shaped: what's the current status, what's still open, what did we just decide. It stops being fine the moment two agents retrieve contradictory "facts" with equal confidence, or a stale definition gets treated as current because nothing distinguished it from something logged five minutes ago. That's the specific mechanism behind the stall this page's subhead points at — not a capability ceiling, a category error, arriving late enough in a rollout that it looks like the technology failed rather than like the architecture never distinguished the four things it needed to.
What follows from here
The next four pages in this section take each memory type in turn — what it should cost to maintain, what governs it, and where organizations already get it wrong even without AI in the loop. After that, a worked example walks a fictional software company's actual artifacts — budgets, incident reports, runbooks, glossaries — through all four tiers at once, including the cases that refuse to sort cleanly into just one. The pages after that take on what a fully assembled architecture looks like, and the harder cross-cutting problems the four-way split doesn't resolve on its own: memory poisoning, multi-agent disagreement, the incentive problem underneath naming a steward, what changes when AI is the one writing the memory rather than just reading it, and what an organization owes the people who ask an agent a question and get an answer back with no sense of which of these four places it actually came from.
Where this framework might not hold
Two limits deserve to be named outright rather than hedged away in every paragraph above. First, this page's own move — reading an organization's memory through a framework built once for individual cognition and once for a single agent's architecture — is itself a further extension, not a third independent discovery. Convergent validity across two prior applications is real evidence the four-way split tracks something genuine; it is not proof that the split survives a third context untouched, and where it doesn't is exactly what the rest of this section is for. Second, Walsh and Ungson's ecological bin — the physical workplace itself as a carrier of institutional knowledge — has no clean equivalent anywhere in this framework, on either the organizational or the agent side, and forcing a fifth category into existence just to keep the map symmetrical would be worse than saying plainly that some kinds of organizational knowledge may not have a software analog at all.
Sources. Jeffrey P. Walsh and Gerardo R. Ungson, "Organizational Memory," Academy of Management Review, Vol. 16, No. 1 (1991), pp. 57–91. Theodore Sumers, Shunyu Yao, Karthik Narasimhan, and Thomas L. Griffiths, "Cognitive Architectures for Language Agents" (2023), arxiv.org/abs/2309.02427. On cognitive memory systems generally: overviews of working, semantic, episodic, and procedural memory from the Open Encyclopedia of Cognitive Science (MIT Press), oecs.mit.edu, and Simply Psychology's summary of long-term memory types, simplypsychology.org. On memory consolidation and Complementary Learning Systems theory: consolidation research summarized in Psychonomic Bulletin & Review, "Examining the relationship between working memory consolidation and long-term consolidation" (2022); CLS theory and 2026 agent-memory survey findings on consolidation as the least-addressed dimension of shipping systems, via Zylos Research, zylos.ai. On Anthropic's Dreaming feature as a production consolidation mechanism: announced at Code with Claude (May 6, 2026); reported detail via the Zylos Research summary above. Park et al., "Generative Agents: Interactive Simulacra of Human Behavior" (Stanford, ACM UIST 2023), as the software-side reflection-and-consolidation precedent, referenced via the Zylos Research summary above.