Overview

The current generation of AI agents is largely stateless: each interaction is treated as an isolated event, with little to no durable memory of past conversations, decisions, or learned preferences. As a result, agents repeatedly ask for the same information, lose context over time, and struggle to improve through experience, especially when interactions are separated by long time gaps.

Agentic Memory addresses this fundamental limitation by introducing persistent, semantic memory as a first-class capability for agentic systems. Instead of relying solely on short-lived context windows or brittle prompt-based workarounds, agents can store meaningful information once and recall it later based on relevance and intent, even months after the original interaction.

By externalizing memory into a managed, long-term store, our Agentic Memory product enables AI systems to behave more like knowledgeable assistants than transactional tools. Agents can accumulate understanding over time, reconnect past knowledge with new situations, and deliver experiences that feel continuous, informed, and increasingly intelligent, regardless of session boundaries or interaction frequency.

Core Concepts

Memory Space

A Memory Space is the top-level container that represents an application, product, or use case. It provides logical isolation and governance boundaries for memory data.

Key characteristics:

• Groups related Memory Stores under a single context
• Scoped to a tenant and business group
• API access is managed at the Memory Space level
• Creation and deletion are managed through the Cloud Portal
• Deleting a Memory Space removes all associated Memory Stores and their memories

Memory Store

A Memory Store is an individual collection of memories within a Memory Space. It is typically scoped to a specific entity such as a user, agent, session, or workflow.

Key characteristics:

• Contains the actual stored memories
• Supports adding, retrieving, listing, searching, and deleting memories
• Can be created and deleted by authorized clients using the Memory Space’s credentials

Memory

A Memory is a single stored text entry.

Key Capabilities

Agentic Memory provides a focused set of high-level capabilities:

• Create and delete Memory Stores within a Memory Space
• Add memories as plain text without managing embeddings manually
• Retrieve memories individually or as a list
• Search memories using semantic similarity against a text query
• Delete memories when they are no longer relevant

All embedding, indexing, and query optimization is handled by the product.

Limits

The default limits are:

• Up to 100 Memory Stores per Memory Space
• Up to 1,000 memories per Memory Store

Custom limits can be configured per instance to support larger-scale or specialized use cases; please reach out to your designated Service Delivery Manager for more information.

Typical Use Cases

1. Personalized AI Assistants

Scenario
Each end user of an AI assistant is assigned a dedicated Memory Store within a Memory Space.

How it works

• A single Memory Space represents the application
• A Memory Store is created for each user
• User-specific preferences, past interactions, and long-term context are stored as memories
• Searches are scoped to the user’s Memory Store only

Benefits

• Strong isolation between users
• Personalized responses based on historical context
• Simple lifecycle management when users are created or deleted
• An ideal pattern for chatbots, copilots, and digital assistants that need long-term personalization.

2. Workflow- or Agent-Centric Memory for Task Automation

Scenario
Memory Stores scoped to workflows, agents, or processes.

Examples

• A customer support automation system uses one Memory Store per support queue or issue type
• An autonomous agent system uses one Memory Store per agent role (e.g., researcher, planner, executor)

How it works

• A Memory Space represents the overall system
• Memory Stores represent shared knowledge for a workflow or agent
• Memories include resolved issues, learned patterns, summaries, or decisions.

Benefits

• Shared institutional knowledge across agents or executions
• Improved consistency and decision-making over time
• An ideal pattern for autonomous agents, RAG pipelines, operational intelligence, and decision-support systems.

Example Use Case: Per-User Memory for a Booking Assistant

Scenario

An AI-powered booking assistant helps users plan and manage travel. Users may interact with the assistant over longer periods that can sometimes span weeks or months, but still expect it to remember their preferences and past decisions.

Each user is assigned a dedicated Memory Store within a shared Memory Space for the application.

What Is Stored as Memory

• Preferred airlines, hotels, or seat types
• Past destinations and recurring travel patterns
• Loyalty program memberships or seating preferences
• Constraints such as budget ranges or preferred travel times
• Notes from previous conversations (e.g. "I prefer aisle seats on long flights")

How It Works Over Time

• On each interaction, the assistant retrieves relevant memories from the user’s Memory Store
• New or updated information is added as additional memories
• Searches are semantic, allowing the assistant to recall relevant context even if phrased differently
• Memory persists across sessions, devices, and long periods of inactivity

User Experience Benefits

• No need to repeat preferences or background information
• More personalized and consistent recommendations
• Conversations feel continuous rather than reset-driven
• Increased trust in the assistant’s usefulness over time

Why Agentic Memory Is a Good Fit

• Strong isolation between users through per-user Memory Stores
• Scales naturally as the user base grows
• Supports long-term personalization without bloated prompts
• Enables agents to improve through accumulated experience

Summary

Agentic Memory provides a low-ops, production-ready foundation for persistent, semantic memory in agentic and AI-driven systems. By abstracting away embeddings, vector storage, and scaling, it allows teams to focus on building intelligent behavior rather than managing infrastructure.