memory

Memory determines how your agent retains, recalls, and reasons over past interactions and knowledge.


Types at a Glance

Type
Persistence
Best For

None

No memory

One-shot lookups, stateless utilities

Conversation

Within session

Interactive chatbots, support agents

Persistent

Across sessions

Personal assistants, user preference tracking

Vector

Semantic retrieval

Large knowledge bases, document Q&A


None

Completely stateless. Every request is independent. The agent has no awareness of previous interactions. Ideal for tools that perform a single action and return a result.

Conversation

Retains a rolling window of recent messages within a single session. When the session ends (user disconnects or TTL expires), memory is cleared.

Settings: Max Messages (default: 50), TTL in minutes (default: 60)

Persistent

Stores facts, preferences, and context permanently across all sessions. The agent remembers everything until entries expire based on TTL.

Settings: Max Messages (default: 50), TTL in minutes (default: 60)

Use cases: An agent that remembers a user's portfolio, risk tolerance, preferred tokens, or notification preferences.

Vector

Stores embeddings in a vector database and retrieves relevant context via semantic similarity search at query time. This is the most powerful memory type — it scales to millions of stored entries.

Settings: Max retrieval results (default: 50), TTL in minutes (default: 60), Vector DB provider and collection name

Use cases: An agent that reasons over hundreds of research papers, a full codebase, or months of market data.


Combining Memory with Data Sources

Vector memory works best when paired with Data Sources and the RAG Pipeline tool. Data sources provide the knowledge; vector memory provides the retrieval mechanism; RAG Pipeline orchestrates the flow.

Was this helpful?