Agent Memory: Short-Term vs Long-Term

As language models evolve into agents that can hold conversations, take actions, or assist across multiple tasks, they need memory — just like humans.

Memory allows an AI agent to:

  • Remember past interactions

  • Track tasks across sessions

  • Personalize responses over time

There are two main types of memory in GenAI systems: short-term and long-term memory.


🧾 1. Short-Term Memory

Short-term memory is the information the model can "see" during a single conversation or session — limited by the context window.

Example:

  • You ask: "Remind me what we discussed earlier?"

  • The model responds accurately as long as it’s within the token limit (e.g., 8,000 or 128,000 tokens)

✅ Used for:

  • Chatbots that remember recent turns

  • Keeping track of the topic during long queries

  • Holding temporary state in tools like LangChain, AutoGen, or ChatGPT

⚠️ Limitation:

  • If the context is too long, older messages are truncated or forgotten


🧠 2. Long-Term Memory

Long-term memory allows the agent to remember facts, preferences, or conversations across different sessions — even after the app is closed and reopened.

This is stored outside the model, typically in:

  • Vector databases (e.g., FAISS, Weaviate)

  • SQL/NoSQL stores

  • Custom memory systems

Example:

  • User says: "My favorite programming language is Python."

  • Weeks later: "Remind me what language I prefer?" → The agent remembers: "You told me Python is your favorite."

✅ Used for:

  • Personalized tutoring or coaching bots

  • AI assistants that evolve with the user

  • Task-tracking agents or multi-session workflows


🔁 Combining Both

Advanced agents blend short-term + long-term memory:

  • Short-term: Active, fast, limited

  • Long-term: Stored, retrievable, structured

Example: A coding assistant remembers what file you were editing last week (long-term) and still tracks your current task (short-term).


📊 Side-by-Side Comparison

Feature
Short-Term Memory
Long-Term Memory

Duration

One session

Across sessions

Stored in

Context window (tokens)

External DB or memory system

Capacity

Limited (8K–128K tokens)

Unlimited (scalable with DB)

Best For

Ongoing conversation flow

Personalization, history, continuity

Example Tool

ChatGPT, LangChain ConversationBuffer

LangChain VectorMemory, AutoGen, Weaviate


🧠 Summary

  • Short-term memory = What the AI can see right now

  • Long-term memory = What the AI remembers over time

  • Smart AI agents use both to feel more natural, helpful, and personal


Last updated