A note on where this comes from. This piece references Citi's Agent Factory org design, discussed at length in the companion piece linked below. As stated there: I have no relationship with Citigroup — financial, employment, consulting, or otherwise — and everything about Citi here is read from its own public job postings and blog post, not from any insider or engagement-based knowledge.
Every organization has a memory problem, and until recently, humans were the only ones who had to solve it. Cognitive science doesn't describe memory as one system — it describes several, each with a different job. Working memory is the transient scratchpad: what you're holding onto right now, for the task in front of you, gone in minutes if nothing reinforces it. Long-term memory is where things go if they survive, and it isn't one bin either. Semantic memory holds general facts and concepts, stripped of the specific moment you learned them — what a word means, what category something belongs to. Episodic memory holds specific events, tagged with when and where they happened — the meeting where the decision got made, not just the decision. Procedural memory holds skills — how to do something, encoded as behavior you can execute without being able to fully narrate it.
These aren't just different shelves for the same kind of thing. They're retrieved differently, updated differently, and — the detail that matters most here — built differently. A semantic fact doesn't usually arrive fully formed; it's abstracted out of episodic memory through consolidation, a real neural process in which a specific experience, retrieved and reinforced enough times across separate contexts, gets generalized into a fact that no longer needs the original episode attached to it. You don't remember the fortieth time you checked that red means stop and green means go. You just know it now.
Organizations already had a name for this
None of this is a novel metaphor. 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 (what specific people happen to recall), culture (how members think, feel, and perceive problems — mostly tacit, absorbed rather than looked up), transformations (the standard processes that turn inputs into outputs, carrying past experience inside the process itself), structures (the org chart and role definitions, which encode who is supposed to know what), and ecology (the physical workplace itself, which holds information nobody wrote down but everybody who's been there picks up).
What's notable, thirty-five years later, is how much of that framework depends on humans as the retrieval mechanism. Culture is retrieved by pattern-matching a colleague's tone. Ecology is retrieved by literally walking through a building and noticing what's where. Individual memory is retrieved by asking Sarah, who's been here longer than the wiki has existed. Every one of those retrieval paths works because humans tolerate ambiguity, infer intent, and know who to ask. That's exactly the capability an AI agent doesn't have by default, which is the actual reason "shared context for AI agents" turned into a live product category in 2026 under names like "Company Brain" — not because organizations didn't have memory before, but because the cheapest available reader of that memory stopped being a person who could shrug and go find out.
Four memories, not one
Bring the four cognitive categories back into a software organization's actual artifacts, and the ProductHarness backplane — as originally defined in When Execution Becomes Abundant — turns out to have quietly merged two of them into a single repository with a single governance rule. This is the gap Three Pillars, No Fourth pointed at without naming: Citi's Agent Factory, Core AI Platform, and Responsible AI functions all need a shared model of firmwide truth, and "shared model of firmwide truth" is doing the work of four different things at once.
Working memory is the decision queue and the in-flight pipeline state: what's currently unresolved, what a fleet coordinator is waiting on, what a work stream's status is this week. It's meant to be cheap to write and short-lived. Most of it should never need to survive past the work stream that produced it, and treating it with backplane-level review would be exactly the wrong amount of ceremony for something this disposable.
Episodic memory is the backplane's "material product decisions with rationale" tier as originally written — a specific, dated record: this is what we decided, on this project, and here's why. It's genuinely useful to the next work stream that hits the same question, but it isn't yet a firmwide fact. It's a precedent, not a law.
Semantic memory is the rest of the backplane's critical tier: the glossary, the entity and state model, the hard constraints. This is knowledge with the originating episode stripped away — nobody needs to know which project first established what "active" means, only that it's settled everywhere now. This is the tier that actually deserves the heaviest governance the original essay specified, because a change here has the widest blast radius of anything in the system.
Procedural memory doesn't live in the backplane at all — it's the production codebase, the reusable component library, the agent skill or template that encodes "how we do this" as executable behavior rather than as a fact you'd read. Getting this wrong doesn't look like a misunderstanding; it looks like a bug, which is why it belongs under engineering review rather than product or content review.
Filing episodic and semantic content under one roof, with one review process, was the original backplane's actual oversimplification. A logged decision and a settled fact aren't the same kind of object, and treating a decision as provisionally true — worth recording, not yet worth enforcing everywhere — versus treating a fact as settled — worth enforcing everywhere, expensive to revisit — are two different jobs that got merged into one.
Consolidation is a decision, not a vibe
The original backplane essay already had an implicit test for what belongs in it: would this be true and relevant to a completely different feature, built by a completely different team, on a completely different product area? That's a consolidation criterion, though the essay didn't name it as one. What the cognitive-science framing adds is the mechanism, not just the test: a memory gets promoted from episodic to semantic through repeated retrieval across independent contexts, not through someone's judgment that it seems generally important.
That distinction is worth taking literally rather than as a nice turn of phrase, because it changes what evidence a promotion decision requires. "This seems like it should apply everywhere" is a guess, made once, by whoever happens to be reviewing the backplane that week. "This specific decision has now been independently retrieved and confirmed relevant in three separate work streams that had no knowledge of each other" is a track record — checkable, loggable, and a much harder thing to be wrong about. The coherence-review function the original essay proposed is well-placed to do this, but it needs to be looking for repetition across independent contexts specifically, not just scanning for anything that looks broadly useful.
The Company Brain market's half-built answer
Worth checking this against what's shipping commercially, because the 2026 "Company Brain" wave is a real, well-funded attempt at exactly this problem, and it's mostly solving a narrower version of it than it advertises. Y Combinator named "Company Brain" one of the ideas it most wants funded in its Summer 2026 Requests for Startups. Meta's internal "AI Second Brain" scaled past 60,000 users across every organizational pillar. The commercial framing describes a knowledge layer combining semantic definitions, entity resolution, governance, lineage, and memory — which, read carefully, is gesturing at the same four-way split this piece is making explicit.
But the typical implementation is one undifferentiated retrieval index — a vector database over Slack, wikis, tickets, and documents — searched with the same mechanism regardless of whether what's being retrieved is a settled fact, a one-off decision, or yesterday's status update. That's functionally an organization that only ever forms episodic traces and never consolidates any of them: everything is technically searchable, and nothing is ever confirmed as generally true, so every retrieval is a fresh similarity-weighted guess instead of a lookup against something the organization actually decided. It's closer to an agent with a very good index of everything it's ever half-overheard than an agent with actual settled knowledge to draw on.
The honest caveat
Two limits worth naming precisely, because they mark exactly where this argument stops being description and starts being a bet. This is a metaphor built on a metaphor: individual cognitive architecture, mapped onto organizations by Walsh and Ungson in 1991, mapped a second time here onto the specific artifacts an AI-agent-consuming organization builds. Two translation layers, two places to check the seams rather than assume they hold.
One seam is visible on inspection: Walsh and Ungson's ecological bin — the physical workplace itself as a carrier of institutional knowledge — has no clean equivalent in a codebase-based system. That's worth stating outright rather than forcing a fifth software-side category into existence just to keep the framework symmetrical. Whatever the ecological bin actually captures — unstated knowledge carried by physical proximity and shared environment — may not have a software analog at all, and saying so plainly is a better answer than pretending otherwise.
The second limit is scale: none of this has been run in production. It's a structural argument, built by taking two established frameworks seriously and applying them somewhere new — a specific, falsifiable claim about how the pieces should fit, not a report of an implementation that exists. Which parts hold under real operating pressure and which quietly re-merge is a question for whoever builds it first, not a question this piece can settle from outside.
The backplane was the right instinct, aimed at the right problem, at the wrong grain size. One governance rule where the underlying analogy, taken seriously, asks for at least four. Any organization building its version of this now doesn't need to decide whether it has a memory problem — it already has one, the way every organization always has, agents or not. What it actually needs to decide is which of the four kinds of memory each new AI agent is being handed, and which kind it's quietly being asked to reconstruct from scratch every single time, because nobody built the version that would let it remember.
Sources. Jeffrey P. Walsh and Gerardo R. Ungson, "Organizational Memory," Academy of Management Review, Vol. 16, No. 1 (1991), pp. 57–91. On memory systems in cognitive science: overviews of working memory, semantic memory, episodic memory, 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 the hippocampus-neocortex relationship between episodic and semantic memory: consolidation research summarized in Psychonomic Bulletin & Review, "Examining the relationship between working memory consolidation and long-term consolidation" (2022). On the 2026 "Company Brain" market: Y Combinator Summer 2026 Requests for Startups; reporting on Meta's internal "AI Second Brain" rollout; Sentra, "What Is a Company Brain? The 2026 Guide," sentra.app; Vectorize, "How to Build a Company Brain for AI Agents," vectorize.io.