A company memory is a living record of what an organisation knows: the decisions it made, the commitments it gave, the context behind both, and how those things changed over time. It differs from an archive in one essential way. An archive stores documents and waits to be searched. A company memory stores connected knowledge and takes part in the work: every AI session reads from it before doing anything else, and writes back what it did.
One idea, many names
The term has cousins, and depending on who is selling to you, you will hear several. They are worth telling apart, because they carry different assumptions about what the thing is and who tends it.
Company brain is the consumer-flavoured version, borrowed from the “second brain” note-taking culture. It usually describes a personal or team tool – notes, bookmarks, saved answers – rather than organisation-wide infrastructure, and it tends to stop at retrieval: a brain you ask, not a memory that takes part in the work.
Organisational memory is the academic term, decades old in management literature. It names the phenomenon – what an organisation retains in its people, processes, and records – rather than any system for keeping it. Useful word, but nobody ever bought one.
Corporate memory is the same idea in enterprise clothing, with roots in records management and knowledge management. In practice it often means the archive – governed, compliant, and searched twice a year.
Institutional knowledge is what people call it when it is walking out of the door. The phrase appears most often in the sentence “we lost a lot of institutional knowledge when she left”, which tells you everything about where it usually lives: in people, unrecorded.
We say company memory because it is the plainest description of the thing we mean: organisation-wide rather than personal, alive rather than archived, and in the flow of work rather than beside it. Underneath all the labels sits the same idea: most of what an organisation knows evaporates, and for the first time the technology exists to stop that happening.
Why the idea is surfacing now
Organisations have always leaked knowledge. Decisions live in someone’s head, context dies in email threads, and every new piece of work starts with the archaeology of finding out what was already known. What changed is the arrival of AI assistants that can actually use knowledge at the moment of work – provided the knowledge is somewhere they can reach, in a form they can reason over.
A folder of PDFs is not that form. Language models reason far better over structured relationships than over piles of flat documents. Give a model a folder of coverage and it summarises; give it a graph of connected decisions, people, projects, and conversations and it can answer the questions people actually ask: what did we decide, why, and what has changed since?
What a company memory is made of
Technically, the substrate is a knowledge graph: people, organisations, projects, meetings, documents, tasks, and the relationships between them. But the graph alone is not the memory, for the same reason a filing cabinet is not a memory. Three more layers make it live.
Automatic capture. Memory that depends on people remembering to file things starves. Meetings should transcribe themselves into structured knowledge; commits, messages, and documents should be swept in from the places where work already happens.
Reasoning with provenance. The most valuable layer captures thinking rather than activity: atomic claims extracted from conversations, each with who said it, where, and when. Disagreements stay alive rather than being flattened, and consensus has to be earned through independent corroboration, not repetition.
An open door for AI. The memory is only useful if every AI session can consult it mid-conversation – in practice, an MCP server exposing the graph to whatever models the organisation uses, with access controlled by role.
This is the architecture we run ourselves and deploy for clients; the longer description lives on the AIP page.
What a company memory is not
It is not a wiki. Wikis hold what someone once wrote down, which drifts from reality the day after it is written. It is not a document management system, because documents are the output of knowledge work, not the knowledge itself: a press release records what you decided to say, not why you said it that way or what the lawyers struck out. And it is not enterprise search. Search finds things; memory knows things, including things that were never written into any single document.
The honest boundary: not everything needs one. If your question is “find me the boilerplate paragraph”, flat search is fine. A memory earns its keep where relationships and time carry the meaning – narratives, stakeholders, decisions that evolve. If a spreadsheet answers your questions, keep the spreadsheet.
What changes when you have one
After a year of running our own: onboarding stops being an oral tradition, because a new colleague’s first week is spent asking the graph rather than interrupting colleagues. Decisions hold, because a choice made in a Monday meeting reaches every AI-assisted piece of work the same afternoon. And the writing improves, because drafting against the memory means citing what the organisation actually did rather than what anyone remembers doing. We wrote up the longer version in what our own company memory looks like after a year.
The caveat we always give: capture is not free. The automatic layers genuinely are automatic, but recording why a decision went the way it did still takes a sentence of discipline in the flow of work. A company memory is an asset you tend, not a system you install. We think that trade is the whole point.
Where to start
Not with an enterprise knowledge programme. Pick the single workflow where connected beats filed most obviously – for communications teams that is usually monitoring, because the data arrives daily and the pain of flat coverage lists is felt every morning. Get one memory-backed briefing into people’s inboxes, let the memory accumulate for a quarter, and the second use case will suggest itself. The compounding is the real argument: everything the memory ingests makes everything else it knows more useful. The longer it runs, the more it knows.
We build company memories for communications, public affairs, and finance teams, and we run our own company on one. If you want to see the machinery, start with AIP; if you want to talk, bring us a problem worth thinking about.