Memory & Context

Memory is what separates a useful agent from a frustrating one. Without it, every conversation starts from zero. With it, an agent can learn your preferences, remember what it worked on yesterday, and build up expertise over time.

The four types of memory

1. In-context memory (working memory)

Everything the model can see right now — your current conversation, any documents you've shared, the agent's recent tool outputs. This lives in the context window and disappears when the session ends.

Analogy: What's currently on your desk.

2. Persistent memory (long-term storage)

Information the agent stores between sessions — facts it's learned about you, past decisions, preferences, knowledge it's built up. Stored in a database and retrieved when relevant.

Analogy: Your notes and files.

3. Episodic memory (history)

A log of past interactions — what tasks were completed, what was decided, what worked. Lets the agent reason about its own past behavior.

Analogy: Your work journal.

4. Semantic memory (knowledge base)

Structured knowledge the agent can search — documentation, company policies, product catalogs, FAQs. Usually stored in a vector database for fast retrieval by meaning, not just keyword.

Analogy: Your reference library.

How retrieval works

Modern agents don't load all their memory into the context at once — that would fill up the context window instantly. Instead they use retrieval — when the agent needs to remember something, it searches its memory store for the most relevant information and loads just that.

User asks question
       ↓
Agent searches memory for relevant context
       ↓
Top results injected into context
       ↓
Model answers with full context

This is why agents can feel like they "remember" things from months ago — they're not storing the full conversation forever, they're storing key facts and retrieving them on demand.

Why this matters for you

When setting up an agent:

• Give it relevant context upfront — paste in your company's style guide, your product catalog, your team's conventions
• Use persistent memory for anything the agent should always know — your name, your preferences, your team's vocabulary
• Be aware of context limits — if a document is too long, the model may not process all of it

Next: Tools & Actions →