What's Semantic Memory?
The distinction that changes everything.
In 1972, cognitive psychologist Endel Tulving proposed two types of memory. This distinction is the foundation for everything we build.
The Distinction That Changes Everything
In 1972, cognitive psychologist Endel Tulving proposed a distinction that would reshape our understanding of human memory. He argued that we don't have one memory system—we have at least two: episodic memory (events) and semantic memory (meaning).
Endel Tulving, 1972Semantic memory remembers meaning, not events, knowledge detached from the episode of learning it. You know that Paris is in France, but you don't remember the moment you learned it. The fact is just... known.
Tulving, 1972Episodic memory is "I remember when I learned this." Semantic memory is "I just know this."
Most AI systems have only episodic memory. They store documents with timestamps. They retrieve based on similarity. They have no concept of what's true—only what was stored.
The Irony of "Semantic" Search
The AI industry is obsessed with 'semantic search,' 'semantic similarity,' 'semantic embeddings.' But most AI memory systems are episodic, not semantic. They use semantic similarity for retrieval while operating as episodic storage. They find things that sound like what you asked for. They don't know what's true.
Most AI systems have only episodic memory. They store documents with timestamps. They retrieve based on similarity to stored records. They have no concept of what's true, only what was stored.
The AI industry is obsessed with "semantic search," "semantic similarity," "semantic embeddings." But most AI memory systems are episodic, not semantic.
They use semantic similarity for retrieval while operating as episodic storage. They find things that sound like what you asked for. They don't know what's true.
RAG doesn't fix this. Retrieval-Augmented Generation finds documents. It doesn't verify claims. It doesn't know which document is current.
It just finds things.
Organizations Don't Have Semantic Memory
Truth fragments when stored in multiple systems. When one source is updated but another isn't, your organization doesn't know the current truth. It has two conflicting episodic records, and no semantic memory to resolve them.
AI can generate unlimited content from your knowledge base. If your knowledge base is wrong, AI generates unlimited wrong content. The verification bottleneck that slowed human authors now paralyzes AI-assisted workflows.
Your organization's knowledge lives in documents. Documents are episodic records—they capture what was written, when it was written, not what's true.
Your sepsis protocol was updated last quarter. The nursing manual wasn't. Both exist. Both get retrieved.
Which one is current?
The system doesn't know. It can't know. Documents don't carry expiration dates.
AI amplifies this problem. AI can generate unlimited content from your knowledge base. If your knowledge base is wrong, AI generates unlimited wrong content. The verification bottleneck that slowed human authors now paralyzes AI-assisted workflows.
Why "Canonical"?
We hold no truths to be self-evident. Every claim an AI reasons from must be explicit, verified, and canonical — not inferred, not assumed, not 'obvious.' The ability to declare what is canonically true is what makes you independent.
SMS Anti-Motto, 2026The term "canonical" is deliberately chosen. From Greek kanōn—a straight rod used by architects as a measuring stick. Later adopted by the Church to describe decrees that were binding on the faithful. Not suggestions. Not interpretations. Settled doctrine that one reasons FROM, not ABOUT.
Why not other terms?
- "Source of truth" is soft. Everything claims to be a source of truth. The term has no teeth.
- "Ground truth" has AI baggage. In machine learning, it means training data. That's not what we mean.
- "Core truth" sounds like opinion. Canonical establishes hierarchy.
When an AI receives a canonical claim, it receives something settled. The AI is not a collaborator on what is true. It is a collaborator on what follows from what has been established as true.
This is the bishop/faithful relationship: bishops set the canon; the faithful reason from it.
The key insight: LLMs don't "hallucinate"—they generate faithfully from an epistemic vacuum. When there's no canon, they produce apocrypha. The solution isn't fixing the model. It's providing bishop-level context.
What Semantic Memory Systems Actually Build
Four principles. One architecture. Systems that remember meaning, not episodes.
The Consciousness Parallel
There's a parallel in consciousness research. Tulving called it noetic consciousness—the awareness that you know something, independent of remembering when you learned it.
Organizations need this shift. From "I remember storing this document" to "I know this claim is true."
From episodic records to semantic knowledge.
This Site Is Built This Way
This site is built this way. The claims on this site aren't scattered across independent pages. They exist as canonical assertions in a structured knowledge base. The pages you see are derived from those claims. When we update a canonical claim, every page that references it can update.
The claims on this site aren't scattered across independent pages. They exist as canonical assertions in canonical/SMS-CLAIMS.json. The pages you see are derived from those claims.
When we update a canonical claim, every page that references it can update. This page demonstrates the methodology it describes.
Who Needs This?
Not every organization needs semantic memory architecture.
- Truth matters (compliance, safety, accuracy)
- You generate multiple outputs from the same knowledge (policies, training, chatbots, FAQs)
- Verification is worth the upfront cost (because downstream maintenance disappears)
- Content is disposable (marketing copy, blog posts)
- Single output (one document, never reused)
- Verification cost exceeds drift cost (rare, but possible)
The Claim Beneath the Claim
Semantic Memory Systems establish canonical truth, verify at the source, generate from verified claims, and stop when they are uncertain. They don't chase perfect recall. They remember what matters.