From knowledge management to AI agent memory.

Before there were vector databases, there was a two-decade argument inside the knowledge-management field about what organizational memory actually is — and most of what AI agent memory is now rediscovering, the KM literature already worked out. It's worth borrowing the answer instead of re-deriving it from scratch.

What organizational memory actually meant

Walsh and Ungson's canonical definition: organizational memory is stored information from an organization's history that can be brought to bear on present decisions. Their real contribution was rejecting the idea of a single repository — memory, in their model, is distributed across individuals, culture, routines, structures, and the organization's external traces. Levitt and March added the behavioral piece: organizations learn by encoding inferences from history into routines, so memory lives partly in people and partly in structures, artifacts, and habits, not only in documents.

Polanyi's distinction between tacit and explicit knowledge sharpens why this matters. Tacit knowledge isn't "missing text" — it's perceptual discrimination, situational judgment, and craft skill that people often can't fully articulate even when asked directly. Nonaka's knowledge-creation theory built an entire organizational model on top of that: firms create knowledge through a spiral of conversion between tacit and explicit forms (socialization, externalization, combination, internalization), and his later concept of ba makes the point explicit — knowledge is inseparable from the context it was created in; strip the context away and it degrades into mere information.

What the 1990s got right

Two complementary instincts emerged, and firms that leaned all-in on one over the other tended to do better than firms that muddled both. Codification — repositories, templates, reusable artifacts — worked for standardized, repeatable problems; Ernst & Young's reuse-focused repository reportedly saved project teams roughly a year of work per engagement. Personalization — communities of practice, apprenticeship, person-to-person transfer — worked for novel, ambiguous, client-specific problems; Bain, BCG, and McKinsey built their model around expert dialogue rather than document reuse. Hansen, Nohria, and Tierney's finding was that trying to do both equally well usually produced confusion and costly infrastructure mismatch, not the best of both worlds.

Buckman Labs and BP both showed the same thing from the personalization side: usable memory depends on trust, networks, and lightweight learning rituals — Peer Assist, After Action Review, Retrospect — not just documents. Brown and Duguid's Xerox research found the real work regularly differs from the documented work, and that meaningful learning happens in the informal communities that form around getting the job done.

Why it stalled

McDermott's diagnosis is still the sharpest: IT made large-scale sharing imaginable, but couldn't deliver KM alone, because most of the knowledge people actually need requires community and interpretation, not just storage. His memorable failure case was the "expensive and useless information junkyard" — a repository that satisfied a documentation mandate while failing every actual worker, because the captured knowledge was too general, too abstract, and too detached from the situation it would be used in.

Fahey and Prusak's related catalog of errors — mistaking data for knowledge, substituting a system for human interface, obsessing over storage metrics instead of whether decisions actually got better — reframes the entire evaluation question from "how much have we stored?" to "what changed in action quality because the memory existed?" March's exploration/exploitation framework explains a subtler failure: systems built around reuse reward what's already legible and familiar, and slowly lose the ability to notice novelty, contradiction, and change. History-dependent memory becomes an asset in one period and a trap in the next.

The direct mapping onto AI agent memory

Modern agent-memory architecture makes the KM lineage explicit rather than metaphorical. CoALA separates working, episodic, semantic, and procedural memory. Generative Agents stores a natural-language experience stream and synthesizes reflections from it. MemGPT treats the context window as fast memory and external storage as an OS-style memory hierarchy. None of this is a coincidence — it's the same design space KM explored at the organizational level, reopened at the agent level.

The KM movement's hardest-won lesson translates directly: usable memory is a socio-technical system, and retrieval matters more than accumulation. A vector store full of embeddings can reproduce every pathology McDermott and Fahey & Prusak diagnosed thirty years ago — weak retrieval relevance, no provenance, no ownership model, no update discipline, no connection back to action. Building an AI memory system without engaging that history means re-discovering the "information junkyard" one production incident at a time.