---
title: "What is semantic memory in AI agents?"
description: "Semantic memory stores an AI agent's general facts: plans, owners, preferences. How facts are distilled from events and kept current as they change."
canonical: https://past.dev/glossary/semantic-memory
last-updated: 2026-09-02
---
# Semantic memory

Source: https://past.dev/glossary/semantic-memory

Semantic memory stores general facts detached from the events that taught them, such as who owns a project or what a customer prefers. An agent consults semantic memory for the current state of the world: plans, owners, preferences, policies. Because the world changes, each fact needs a validity window and a link to what replaced it, or the memory drifts out of date.

## How semantic memory is built

Semantic facts are distilled from events. A sales call on May 3 in which Acme Corp agrees to upgrade produces a fact: Acme is on the Enterprise plan. The call is episodic; the fact it establishes is semantic. The step that turns raw text into such facts is [fact extraction](/glossary/fact-extraction). The term itself comes from cognitive psychology, where semantic memory names general knowledge separated from the experiences that produced it.

A well-formed semantic fact keeps four parts:

- **Subject**: the resolved entity the fact is about. See [entity resolution](/glossary/entity-resolution).
- **Value**: the content, such as the plan name or the project owner.
- **Validity**: when the fact became true and when it stopped being true. See [validity window](/glossary/validity-window).
- **Evidence**: the events that established the fact, so it can be checked.

## Why it matters for AI agents

Semantic memory is what an agent consults for the current state of the world: plans, owners, budgets, preferences, policies. Because the world changes, a semantic store without update handling serves stale answers. A [knowledge update](/glossary/knowledge-update) must supersede the old value rather than overwrite it, so the current value is returned by default and the history stays queryable. See [fact supersession](/glossary/fact-supersession).

How fact recall is evaluated against a stored history is described in the [benchmark methodology](/benchmarks/methodology). [past.dev](/docs/memory-api/quickstart) extracts facts from ingested text and tracks each one over time, with the previous value and its dates kept alongside the current one.

## Semantic memory in practice

A support agent opens a ticket from Acme Corp. Semantic memory supplies the standing facts: Acme is on the Enterprise plan, effective May 3. The account owner is Priya Shah. Invoices go to Acme's billing alias. None of these is an event. Each was established by events and remains true until superseded.

When Acme downgrades on November 20, the plan fact is replaced. The old fact keeps its validity window, May 3 to November 20, and the new fact starts its own. An agent asked in December answers with the new plan; an agent asked about October answers with the old one. See [point-in-time recall](/glossary/point-in-time-recall).

## Related concepts

- **[Episodic memory](/glossary/episodic-memory)**: the specific events facts are distilled from, such as the May 3 call.
- **[Procedural memory](/glossary/procedural-memory)**: how tasks are done rather than what is true.
- **[Working memory](/glossary/working-memory)**: the facts currently loaded into the context window.
- **[Knowledge update](/glossary/knowledge-update)**: the controlled replacement of a fact's value when the world changes.
- **[Memory staleness](/glossary/memory-staleness)**: what accumulates when semantic facts are never updated.

## Frequently asked questions

### What is semantic memory in AI?

Semantic memory is the store of general facts an AI agent knows, such as a customer's plan or a project's owner. The facts are kept separate from the specific events that established them.

### How is semantic memory different from episodic memory?

Semantic memory holds general facts, such as which plan a customer is on. Episodic memory holds the individual events, such as the call in which the customer chose that plan.

### How does an AI agent keep semantic memory up to date?

Each fact carries the time it applies to, and a newer fact supersedes the older one rather than overwriting it. The current value is returned by default and the history stays queryable.

## Related

- [Episodic memory](https://past.dev/glossary/episodic-memory)
- [Knowledge update](https://past.dev/glossary/knowledge-update)
- [Fact supersession](https://past.dev/glossary/fact-supersession)
- [Facts that change over time](https://past.dev/guides/facts-that-change-over-time)
- [Memory API quickstart](https://past.dev/docs/memory-api/quickstart)