Skip to content

01-short-term-vs-long-term-memory.md

Latest commit

 

History

History
78 lines (54 loc) · 3.27 KB

01-short-term-vs-long-term-memory.md

File metadata and controls

78 lines (54 loc) · 3.27 KB
Raw

Short-term vs. Long-term Memory

Agent Memory is the ability of an AI agent to store, update, and recall information from past interactions to guide current and future decisions.

Without memory, an agent starts every session with a blank slate, repeating questions and making the same mistakes.

Experience ──► Store in Memory ──► Retrieve when Relevant ──► Inject to Context ──► Action

Short-Term vs. Long-Term Memory

Feature Short-Term Memory (Working Memory) Long-Term Memory (Persistent Memory)
Lifespan Temporary (Current session / execution only) Permanent (Persists across sessions and days)
Storage Location In-Context prompt & RAM External Database, Vector DB, or Key-Value store
Purpose Track immediate task steps, tools, and variables Remember user preferences, past executions, and facts
Capacity Limited by context window size Virtually unlimited

1. Short-Term Memory (Working Memory)

Short-term memory holds the information the agent needs right now to complete its current task:

  • Current user messages and conversation turn.
  • Intermediate tool outputs and scratchpad reasoning.
  • Active variables and workflow state.

Once the session ends or the agent resets, short-term memory is cleared.

2. Long-Term Memory (Persistent Memory)

Long-term memory stores distilled experiences outside the context window and retrieves them when relevant:

  • Semantic Memory: Facts and user profile (e.g., "User uses macOS and Python 3.12").
  • Episodic Memory: Specific past experiences (e.g., "Last week we fixed a Docker build error in module X").
  • Procedural Memory: Rules and instructions (e.g., "Always run linter before committing").

Agent Memory Architecture

Minimal Example (Working vs. Persistent Memory)

class AgentMemory:
    def __init__(self, user_id: str, db_client):
        # Short-term memory (in-memory for current run)
        self.working_memory = []
        
        # Long-term memory (persistent store)
        self.user_id = user_id
        self.db = db_client

    def remember_fact(self, fact: str):
        """Save to persistent long-term storage."""
        self.db.save_memory(user_id=self.user_id, text=fact)

    def load_relevant_memories(self, query: str) -> list[str]:
        """Retrieve relevant past memories into active context."""
        return self.db.semantic_search(user_id=self.user_id, query=query, top_k=2)

Short-term memory is what the agent is thinking about right now; long-term memory is what the agent has learned over time.

References & Further Reading