Agent Memory API

How it works

Instead of manually managing vector stores, relational state, and LLM context limits, Pulsbase natively manages tiered memory for your agents:

• Working Memory: Fast, relational storage for the last N messages of a conversation.
• Archival Semantic Memory: When working memory fills up, a background edge model summarizes the oldest messages, embeds the summary, and stores it in vector search.
• Entity Extraction: Pulsbase actively extracts facts and entities into a Temporal Knowledge Graph in the background.

Using the SDK

Everything is handled through the single @pulsbase/client SDK.

// 1. Add a new message to a session
await db.memory.add('session_abc123', {
  role: 'user',
  content: 'Remind me to use Python for the backend.'
});

// Pulsbase automatically manages context limits in the background!

// 2. Fetch perfectly optimized context for your LLM prompt
const context = await db.memory.getContext('session_abc123');

/*
context returns:
{
  working_memory: [...], // Last 10 raw messages
  archival_context: [...], // Semantically relevant summaries
  entities: {...} // Extracted facts
}
*/

Pricing: Memory Syncs

We don't drain your generic AI Neurons silently in the background. Instead, we use a transparent, predictable metric called Memory Syncs.

• What is a Memory Sync? One sync equals the background pipeline that summarizes an overflowed working memory, extracts entities, and generates vector embeddings.
• Hobby Plan: Includes 5,000 free Memory Syncs per month.
• Pro Plan: Includes 50,000 Memory Syncs, and $0.002 per additional sync.