Agent Memory Types: Episodic, Semantic & Working Memory in AI Agents

A truly capable agent doesn't just respond to the current message — it remembers what happened before, knows what it's learned, and can retrieve relevant past context. This requires implementing memory deliberately, because LLMs themselves are stateless between API calls.

The three memory types

🧠

Working Memory (In-Context)

Everything in the current context window — the conversation so far, system prompt, retrieved chunks, and tool outputs. Fast to access (already in context) but limited by context window size and resets every session.

Immediate contextThis conversationTool outputs

📖

Episodic Memory (Conversation History)

Stored records of past interactions — summarized or verbatim. Lets the agent recall "last week you mentioned you prefer Python" or "we discussed this issue in March." Stored externally (database) and retrieved selectively into context.

Past conversationsUser preferencesPrior decisions

🗃️

Semantic Memory (Long-Term Knowledge)

Domain knowledge embedded and stored in a vector database. The agent retrieves the most relevant knowledge chunks at query time — product documentation, policies, past solutions. This is the foundation of RAG.

DocumentsKnowledge basesPoliciesVector search

Combining all three with MoltBot

from moltbot import Agent
from moltbot.memory import EpisodicMemory, SemanticMemory

agent = Agent(
    model="claude-sonnet-4",
    memory=EpisodicMemory(
        store="postgres",            # persist conversations
        max_history_tokens=2000,     # how much history to inject
        summarize_after=20,          # auto-summarize long histories
    ),
    knowledge=SemanticMemory(
        vector_store="pinecone",      # your knowledge base
        top_k=5,                     # chunks per query
    ),
)

response = agent.chat(
    user_id="user_123",             # scopes episodic memory per user
    message="What did we discuss about pricing?"
)
      

When to use each type

Working memory only — Simple Q&A bots, single-turn tasks where session history doesn't matter.
+ Episodic — Customer service agents, personal assistants, any use case where remembering the user's history improves quality.
+ Semantic — Knowledge-intensive tasks: support bots, internal wikis, research assistants. Anywhere the agent needs to recall facts beyond what fits in context.
All three — Enterprise agents that need to know both the user's history AND the company's knowledge base simultaneously.

All three memory types on MoltBot

Episodic, semantic, and working memory — configured per agent. 14-day free trial.

Start Free Trial →