Agent memory: how AI systems remember, and why most forget everything

Workingthis session onlyminutes

Episodicwhat happenedsessions

Semanticwhat we knowmonths

Proceduralhow we do thingspermanent

In plain terms

Think about the difference between two people helping you. The first is a contractor you re-brief from scratch every single morning — capable, but permanently new. The second is a colleague in year three: she remembers what you tried in March, knows the codebase’s odd corners, and has absorbed how your team likes things done. You never explain twice.

Most AI tools today are the contractor. Close the session and everything is gone — the conventions it learned, the dead ends it explored, the decisions you made together. Tomorrow you pay to teach it again. The systems worth building — and buying — are working on becoming the colleague.

The anatomy

The useful taxonomy comes from cognitive science and maps cleanly onto engineering. Working memory is the context window — fast, small, and wiped between sessions (Episode 3 covered managing it). Episodic memory is the record of what happened: past sessions, past attempts, past decisions — typically stored as searchable logs and summaries, retrieved when relevant. Semantic memory is what the system knows about your world: the architecture of your codebase, your domain’s vocabulary, the constraints that keep recurring — accumulated and stored outside any session. Procedural memory is the most valuable and least common: how things should be done — your standards, your patterns, your definition of good — encoded so the agent applies them without being told each time. (That one gets its own episode next.)

The honest trade-offs: memory can go stale (yesterday’s fact, today’s bug), retrieval can surface the wrong memory at the wrong time, and remembering everything raises real privacy and access questions. Memory is a feature you engineer and govern — not a switch you flip.

What good looks like

One question cuts through vendor decks: does this system’s knowledge of our codebase compound or reset? Ask what it remembers a week later, where that memory lives, who can see it, and how staleness is handled. “It has memory” is not an answer; those four are.

Where Cortex stands

This is what we mean by delivery intelligence in Cortex: what the platform learns about your codebase and your standards persists, so each cycle starts further ahead than the last. And the Kinetic Library is procedural memory at the company scale — fourteen years of how-things-should-be-done, encoded once, applied everywhere.

Questions this paper answers

What are the types of AI agent memory? Working (the current context), episodic (records of past sessions), semantic (accumulated facts about your domain), and procedural (encoded ways of doing things). Production systems build the last three deliberately; the first comes free and vanishes.

Why does my AI tool forget everything between sessions? It only has working memory — the context window — which is cleared when the session ends. Persistence requires engineered episodic, semantic, and procedural memory outside the window.

How can an AI agent remember our codebase? By writing what it learns — structure, conventions, decisions — to external storage as it works, and retrieving it in future sessions. The test is whether its usefulness compounds over weeks instead of resetting daily.