What is semantic memory?

Episodic memory records what happened, tied to a moment and a sequence of events. A great deal of what an agent needs to know isn’t like that at all, it’s simply true, generally, independent of when or how it came up. That’s semantic memory: facts and knowledge that hold regardless of the specific occasion that revealed them. It behaves differently from episodic memory in almost every practical respect, from how it gets consolidated to how much of it should exist at all, and it’s worth understanding those differences directly rather than treating “long-term memory” as one undifferentiated pile.

What Actually Qualifies as Semantic Rather Than Episodic?

A semantic memory states something as generally true, disconnected from any particular occasion: “the user is a vegetarian,” “this client’s account requires two-factor approval for wire transfers over ten thousand dollars,” “the company’s fiscal year starts in April.” None of these describe an event that happened at a specific time. Each one describes a standing state of the world that should hold true whenever it gets checked, regardless of which conversation happened to reveal it first.

The test is straightforward: would this statement still make sense with no date or occasion attached to it at all? “The user is a vegetarian” reads perfectly fine on its own, with no reference to when or how that came up. “The user asked about vegetarian options during a frustrating call about a mixed-up order” doesn’t read the same way stripped of its context, because it’s actually describing something that happened, not a standing fact about the world.

Why Does the Same Semantic Fact Getting Mentioned Repeatedly Actually Cause a Problem?

Episodic memory is comfortable holding many separate records of similar things happening on different occasions, because each occurrence really is its own distinct event. Semantic memory has no equivalent tolerance. If someone mentions being vegetarian in ten different conversations, phrased ten different ways, that should collapse into exactly one canonical fact, not ten near-duplicate entries competing with each other every time something food-related gets searched for.

This isn’t just good hygiene, it’s the actual defining requirement for something to function as semantic memory at all. A pile of ten unresolved restatements of the same fact isn’t “the fact, stored ten times for safety.” It’s an unresolved mess that forces whatever retrieves it to figure out, again, that they’re all saying the same thing, exactly the kind of repeated, avoidable work covered earlier as one of memory’s core failure modes.

Does Semantic Memory Ever Need to Change Once Established?

Being timeless in form doesn’t mean being permanent in content. People change diets, companies revise policies, circumstances shift in ways that make a previously accurate fact flatly wrong going forward. The difference from episodic memory isn’t that semantic facts never change, it’s how they age: an old event gradually matters less as time passes, while a semantic fact stays exactly as relevant right up until the specific moment it becomes incorrect, with no gradual fading in between.

This means semantic memory needs correction in place rather than gradual decay. When a fact changes, the old version isn’t simply less useful now, it’s actively wrong, and needs to be amended directly rather than just supplemented with a newer, competing entry sitting alongside the outdated one.

Is There Usually Less Semantic Memory Than Episodic Memory?

Generally, yes, and that’s a sign of correct consolidation rather than a limitation. Semantic memory is meant to be dense and canonical: a relatively small number of durable, broadly applicable statements, each one load-bearing. Episodic memory naturally accumulates far more entries over the same stretch of time, since every distinct occurrence earns its own record rather than collapsing into something else.

A well-maintained semantic memory store for a single ongoing relationship might amount to a few dozen durable facts even after years of continued use, while the episodic record for that same relationship could run into the thousands. That imbalance isn’t a red flag. It’s exactly what correct behavior looks like for these two categories, since semantic memory is doing its job precisely by staying small, consolidated, and canonical rather than growing in proportion to how much has been said.

How Does Weaviate Engram Keep Semantic Memory Consolidated and Current?

Weaviate Engram treats new semantic-shaped information the way it needs to be treated: checked against what’s already known and reconciled into a single, current statement rather than appended as one more entry in a growing pile. Consider a corporate travel-booking assistant for employees, where a fact like an employee’s home office location needs to stay canonical and get corrected the moment it changes:

from engram import EngramClient

client = EngramClient(api_key=os.environ["ENGRAM_API_KEY"])

client.memories.add(
    "I just relocated to the Austin office, so future trips should be booked from there instead of Chicago.",
    user_id="employee-8834",
)

The earlier fact about a Chicago home base doesn’t sit alongside this new one as two competing statements. It gets reconciled the moment the update arrives, so a later search returns exactly one current answer:

result = client.memories.search(
    query="Which office is this employee based out of for booking purposes?",
    user_id="employee-8834",
)

Whether this employee relocated once or several times over the years, the semantic record stays exactly as compact as it should be: one current fact, correctly updated in place, never a pile of superseded office locations left for something downstream to sort through. That’s the whole discipline semantic memory requires, holding not more information but the right, current version of it, however many times the underlying reality has changed since it was first recorded.

Knowing that something is true, like a preference or a standing fact, is different from knowing how to actually do something well, a skill or a workflow refined through repeated experience. Our next chapter, What is procedural memory?, turns to exactly that third category.