Agent Memory: Short-Term, Long-Term & Episodic Memory for AI Agents

Memory is one of the least-discussed yet most impactful dimensions of agent design. Without memory, every conversation starts cold. With well-designed memory, agents accumulate knowledge, remember user preferences, and build on prior work — much closer to how a skilled human assistant operates.

The three types of agent memory

⚡

Short-Term (In-Context) Memory

Everything currently in the model's context window — conversation history, tool results, documents loaded for this session. Fast, perfect recall, zero setup. But limited to the context window size and wiped when the session ends.

EphemeralPerfect recallContext-limitedZero setup

🗃️

Long-Term Memory (Vector / Key-Value Store)

Facts, preferences, and summaries persisted to an external store — retrieved via semantic search (vector DB) or exact lookup (key-value store). Survives across sessions, scales to millions of entries. Requires a retrieval step introducing latency and imperfect recall (~80–95%).

PersistentScalableRetrieval latency~90% recall

📔

Episodic Memory

Complete records of past sessions, interactions, or task executions stored as structured logs. Lets the agent reference "what happened last time" or learn from prior failures. Most powerful for agents working iteratively on long-horizon tasks.

Task historyLearn from pastStructured logs

MoltBot memory configuration

from moltbot import Agent, VectorMemory

agent = Agent(
    model="claude-opus-4",
    memory=VectorMemory(
        store="pinecone",             # or "weaviate", "pgvector"
        namespace="user-{user_id}",    # per-user isolation
        auto_summarize=True,           # compress old messages
        ttl_days=90                      # auto-expire old memories
    )
)
# Explicit write
agent.memory.store("User prefers bullet-point summaries over prose")
# Memories retrieved automatically at query time
response = agent.run("Summarize this week's metrics")
      

Choosing the right architecture

In-context only: Fine for single-session tasks with no history needs. Simple, but the agent can't grow over time.
Vector long-term memory: Best for personal assistants, CRM agents, any use case where preferences accumulate over time.
Episodic + in-context: Best for agents iterating over multi-day tasks. Episodic log lets the agent resume exactly where it left off.
Full hybrid (episodic + vector): The production architecture for complex enterprise agents — highest capability, most implementation effort.

Native memory for production agents on MoltBot

Vector memory, episodic logs, per-user stores. 14-day free trial.

Start Free Trial →