Vector Search in Streaming Databases: Real-Time Similarity at Scale

AI agents need persistent memory — a record of past interactions, current context, and accumulated knowledge that persists across conversations. A streaming database serves as the agent memory layer: materialized views that continuously aggregate agent interactions into queryable state.

Agent Memory Architecture

User interactions ──→ Kafka ──→ RisingWave
                                    │
                         ┌──────────┴──────────┐
                         ↓                      ↓
                   Short-term memory      Long-term memory
                   (session context MV)   (historical aggregation MV)
                         │                      │
                         └──────────┬──────────┘
                                    ↓
                              Agent queries via PG

Implementation

-- Short-term: current session context
CREATE MATERIALIZED VIEW agent_session AS
SELECT user_id, session_id,
  array_agg(message ORDER BY ts) as conversation_history,
  last_value(intent ORDER BY ts) as current_intent,
  COUNT(*) as turns
FROM agent_interactions
WHERE ts > NOW() - INTERVAL '30 minutes'
GROUP BY user_id, session_id;

-- Long-term: user preferences and history
CREATE MATERIALIZED VIEW agent_memory AS
SELECT user_id,
  COUNT(DISTINCT session_id) as total_sessions,
  array_agg(DISTINCT topic) as discussed_topics,
  last_value(sentiment ORDER BY ts) as last_sentiment
FROM agent_interactions GROUP BY user_id;

Frequently Asked Questions

Why not store agent memory in PostgreSQL?

PostgreSQL stores static data. A streaming database continuously processes interaction events and maintains pre-aggregated memory views in real time — no batch jobs needed to update memory.

How does this compare to LangChain memory?

LangChain memory stores conversation history in-process or in Redis. Streaming database memory persists across sessions, aggregates across interactions, and is shared across agent instances.

AI Agent Memory: How Streaming Databases Power Agent State