agno-memory-sessions

Attach a Db backend (DynamoDB or SQLite) to the Agent and set update_memory_on_run=True with add_memories_to_context=True. Always pass a user_id so factual memories are stored per user and persist across sessions.

Sessions store the message history of a single conversation, keyed by session_id. User memories store durable facts about a user, keyed by user_id, and are retrieved across different sessions to personalize responses.

Use update_memory_on_run for predictable cost — it extracts facts once per run. enable_agentic_memory gives the agent autonomy but triggers extra LLM calls per memory operation, so reserve it for high-value, long-term relationships where recall quality justifies token spend.

Use a cheap dedicated model in the MemoryManager (e.g., gpt-4o-mini), prefer automatic mode over agentic mode, limit retrieval with last_n and a low limit, and periodically prune old memories. These steps reduce both extraction cost and context size.

Agno does not expose a native TTL parameter on its DynamoDb class. You must enable TTL on the table yourself (e.g., via update-time-to-live with an expires_at attribute) and ensure items carry the expiration epoch attribute, since DynamoDB only expires items containing it.

Yes, Agno provides SqliteDb for local development without AWS, requiring sqlalchemy. Switch backends via an environment variable so the same agent code runs locally on SQLite and in production on DynamoDB.