Memory Architecture

Add Persistent Memory to Vertex AI

Give your Vertex AI agent a brain. Smriti adds long-term, causal memory — retaining context across every session, understanding why events happened, not just what happened.

Read the Docs Back to Home
O(1)
Connection Retrieval
S-V-O
Extraction Engine
Dual
pgvector Architecture
∞
Context Retention
The Problem

Vertex AI agents lose all context when a session ends. Every conversation starts from zero. Users repeat themselves, agents forget preferences, and causal connections between events disappear entirely.

The Smriti Solution

Smriti integrates directly with Vertex AI as a drop-in memory layer. It extracts Subject-Verb-Object triples from every interaction, stores them in a dual semantic and episodic pgvector space, and retrieves causal context in O(1) time — across any number of sessions.

How It Works with Vertex AI

Vertex AI calls Smriti's memory API before generating any response. Smriti checks its timeline, retrieves relevant causal context, and passes it as structured memory to the Vertex AI prompt. The agent responds with full awareness of past events — not just the last message.

# Install Smriti pip install smriti # Initialize with vertex-ai from smriti import SmritiMemory memory = SmritiMemory( backend="pgvector", mode="dual", # semantic + episodic svo_extraction=True # auto S-V-O tagging ) # Store any interaction memory.store("User bought flowers on Tuesday") # Retrieve with causal context result = memory.recall("why did we argue on Friday?") # Returns: Tuesday → flowers → context → Friday

Smriti vs Standard Vector DB

Feature Standard Vector DB Smriti
Long-term memory Partial Full timeline
Causal context (A caused B) No S-V-O graph
Cross-session retention Session only Persistent
Encrypted memory blocks No Block-level
Query speed O(n) scan O(1) via index

Why Vertex AI needs long-term memory

Vertex AI is a powerful AI framework but lacks persistent, cross-session memory by default. Each agent run begins with no knowledge of previous interactions. For production applications — customer support, personal assistants, research tools — this is a critical limitation. Smriti solves this by maintaining a chronological memory timeline that persists indefinitely.

How Smriti's S-V-O architecture improves Vertex AI agents

Unlike simple vector search that retrieves semantically similar text, Smriti extracts the Subject-Verb-Object structure from every stored memory. This means a Vertex AI agent can answer causal questions: not just 'what happened?' but 'why did it happen?' and 'what was the consequence?' This is the difference between a database and a timeline.

Getting started with Smriti and Vertex AI

Smriti installs as a Python package and integrates with Vertex AI in under 10 lines of code. It requires a PostgreSQL instance with pgvector enabled. Memory is stored in two parallel vector spaces — one for semantic meaning, one for episodic cause-and-effect — giving Vertex AI agents both broad recall and precise causal reasoning.