How do AI Agents manage Short-Term, Long-Term, and Episodic Memory?
❓ Interview Prompt: Compare short-term context window memory, long-term vector DB memory, and episodic session reflection.
Autonomous AI Agents require memory systems to maintain context across multi-turn user interactions:
1. Short-Term Memory:
Stored directly inside the LLM Context Window. Maintains recent conversation history and active tool results. Pruned or summarized when token limits are reached.
2. Long-Term Memory:
Stored in Vector Databases or Key-Value stores. Holds persistent facts, user preferences, and domain knowledge retrieved via semantic search.
3. Episodic / Working Memory:
Stores structured logs of past agent problem-solving sessions. Allows agents to reflect on previous mistakes and reuse proven action plans.
Code Implementation Example
class AgentMemoryManager {
private shortTermContext: Message[] = [];
private longTermVectorStore: VectorStore;
async rememberUserPreference(userId: string, preference: string) {
const vector = await embedText(preference);
await this.longTermVectorStore.upsert({ id: userId, vector, metadata: { preference } });
}
async recallPreferences(userId: string, query: string): Promise<string[]> {
const queryVec = await embedText(query);
const results = await this.longTermVectorStore.query({ vector: queryVec, filter: { userId } });
return results.map(r => r.metadata.preference);
}
}Agents combine short-term context window memory with long-term vector store persistence and episodic session reflection.