Every agent session starts from scratch. That has been the reality for most AI agents today. Ask an agent to help you resolve a billing dispute, and it might ask you the same clarifying questions it asked last week, with no recollection of your preferences, your previous decisions, or the work you’ve already done together. This is not a limitation of the LLM, it is a fundamental gap in how agents are built. Without persistent memory, every interaction starts at zero.
Enterprises deploying AI agents face a challenge that is easy to overlook but costly in practice: agents are stateless by default. When a session ends, the context ends with it. Preferences, past decisions, in-progress workflows, and the outcomes of previous conversations all disappear.
This creates friction in exactly the scenarios where agents are supposed to save time, for example:
A collection agent asks a customer to re-confirm payment preferences it has already captured.
An HR assistant forgets that a user prefers documents in German and generates English responses each time.
A configuration transport agent has no memory of approvals completed in a previous session, restarting an hours-long workflow from the beginning after a timeout.
Stuffing chat history into the prompt as a workaround is expensive, fragile, and hits context window limits fast. What agents need is not a bigger prompt, they need real, durable memory.
Introducing SAP Agent Memory Service Powered by SAP HANA Cloud
SAP Agent Memory Service is a fully managed, persistent memory layer for AI agents on SAP BTP. It gives agents the ability to store knowledge across sessions, retrieve it semantically when it is needed, and maintain an auditable log of their interactions, all backed by SAP HANA Cloud’s vector engine and governed with enterprise-grade retention controls.
Powered by SAP HANA Cloud’s built-in NLP capabilities, Agent Memory Service is not just a key-value cache. It is an evolving memory store that allows agents to write what they learn, recall what matters by meaning rather than keyword, and improve over time as memories accumulate. All agent memories and messages are stored directly in your SAP HANA Cloud instance, the same trusted, enterprise-grade database that already powers your business data. so, your agent context lives where your data lives, governed by the same security and compliance controls you already rely on.
SAP Agent memory Service Powered by SAP HANA Cloud (Note – *Planned items 2027)
How It Works
Agent Memory Service stores two types of records:
Memories are durable, semantically searchable knowledge entries – facts, decisions, summaries, and preferences that an agent should recall across future sessions. When an agent writes a memory, the service automatically generates a vector embedding using SAP HANA Cloud’s NLP engine. At retrieval time, a vector similarity search surfaces the most relevant memories for the current context, even when the query shares no keywords with the stored content.
Messages are raw, chronological conversation logs grouped by session. They give agents an auditable, replayable record of every interaction – essential for compliance, handovers, and multi-turn context within an active session.
All data is scoped to an (agentID, invokerID) pair, ensuring that memories and messages are isolated per agent and per user, with no cross-contamination between tenants, agents, or users.
The service exposes a standard OData REST API with four key endpoints:
No proprietary SDK is required. An optional Python client is available via the SAP Cloud SDK for Python for those who prefer it.
What Becomes Possible
Agent Memory Service transforms what an agent can do across three dimensions:
Continuity across sessions – A user’s preferences, in-progress tasks, and past decisions survive session boundaries. When a user returns, whether hours or weeks later, the agent picks up where they left off. For example – “Customer prefers consolidated invoices” is automatically recalled in every billing cycle, without the user needing to say it again.
Leaner, more precise prompts – Rather than appending entire conversation histories to every prompt, agents retrieve only the top-K semantically relevant memories per turn. This keeps prompts lean, reduces LLM costs, and improves response quality.
Improvement over time – As memories accumulate and access timestamps update, agents get sharper without retraining. Frequent dispute patterns are recognized and resolved faster. Recurring questions are anticipated and addressed proactively.
Human oversight and control – Users can inspect, edit, or delete their memories directly, providing full transparency. Audit teams can trace how a billing decision or dispute was handled, turn by turn, using the message log.
Getting Started
Agent Memory Service is available as a standalone SAP BTP service subscription tied to your SAP HANA Cloud instance (Bring Your Own HANA plan).
Provision via SAP BTP Service Marketplace – In your SAP BTP subaccount, add the hana plan for SAP HANA Cloud and the hana-agent-memory plan as entitlements. Provision an SAP HANA Cloud instance or use your already existing one (make sure to enable “nlp”: true in the ‘Advanced Setting’ configuration – this is required for semantic search). Then create an Agent Memory Service instance using the ‘hana-agent-memory’ plan, link it to your HANA Cloud instance, and create a service key. The key gives you the clientid, clientsecret, url, and application_url fields needed to authenticate via OAuth 2.0 client credentials.
Subscribe to ‘hana-agent-memory’ service plan on SAP BTP Cockpit
Call the API directly or use the SAP Cloud SDK for Python – Once provisioned, you can call the OData REST API from any HTTP client. For Python-based agent development, the SAP Cloud SDK for Python provides an optional client that handles authentication and request handling, with ready-to-use patterns for storing memories, running semantic search, logging conversation turns, and managing retention.
The journey toward agentic SAP HANA Cloud has been about making agents smarter at every layer, from the data they discover and query, to the context they carry with them over time.
Agent Memory Service addresses the most fundamental limitation in today’s agent architectures: the absence of persistent, governed, enterprise-grade memory. With it, agents stop being one-shot tools and become evolving partners.
For developers, this means:
A managed, persistent memory layer that works across SAP and non-SAP agent frameworks (LangChain, LangGraph, SAP BAIP, and custom agents)
Semantic retrieval powered by SAP HANA Cloud’s vector engine, no external embedding service required
Native integration in Joule Studio (planned towards 2027) and the SAP Cloud SDK for Python
Built-in governance: configurable retention, data isolation, OAuth 2.0 security, and audit logging
With Agent Memory Service now generally available, SAP HANA Cloud continues to establish itself as the database for SAP’s AI-native software architecture, not just storing data, but helping agents remember, learn, and improve across every interaction.
Note on availability: The service is currently available on a bring-your-own SAP HANA Cloud model, where you provision and manage your own HANA Cloud instance. A fully managed offering, where SAP handles the underlying infrastructure entirely, with no HANA Cloud instance required on your end, is planned towards the end of 2026. Stay tuned for updates.
Every agent session starts from scratch. That has been the reality for most AI agents today. Ask an agent to help you resolve a billing dispute, and it might ask you the same clarifying questions it asked last week, with no recollection of your preferences, your previous decisions, or the work you’ve already done together. This is not a limitation of the LLM, it is a fundamental gap in how agents are built. Without persistent memory, every interaction starts at zero.With the General Availability of SAP Agent Memory Service in Q3 2026, that changes.The Problem: AI Agents That Forget EverythingEnterprises deploying AI agents face a challenge that is easy to overlook but costly in practice: agents are stateless by default. When a session ends, the context ends with it. Preferences, past decisions, in-progress workflows, and the outcomes of previous conversations all disappear.This creates friction in exactly the scenarios where agents are supposed to save time, for example:A collection agent asks a customer to re-confirm payment preferences it has already captured.An HR assistant forgets that a user prefers documents in German and generates English responses each time.A configuration transport agent has no memory of approvals completed in a previous session, restarting an hours-long workflow from the beginning after a timeout.Stuffing chat history into the prompt as a workaround is expensive, fragile, and hits context window limits fast. What agents need is not a bigger prompt, they need real, durable memory.Introducing SAP Agent Memory Service Powered by SAP HANA CloudSAP Agent Memory Service is a fully managed, persistent memory layer for AI agents on SAP BTP. It gives agents the ability to store knowledge across sessions, retrieve it semantically when it is needed, and maintain an auditable log of their interactions, all backed by SAP HANA Cloud’s vector engine and governed with enterprise-grade retention controls.Powered by SAP HANA Cloud’s built-in NLP capabilities, Agent Memory Service is not just a key-value cache. It is an evolving memory store that allows agents to write what they learn, recall what matters by meaning rather than keyword, and improve over time as memories accumulate. All agent memories and messages are stored directly in your SAP HANA Cloud instance, the same trusted, enterprise-grade database that already powers your business data. so, your agent context lives where your data lives, governed by the same security and compliance controls you already rely on.SAP Agent memory Service Powered by SAP HANA Cloud (Note – *Planned items 2027)How It WorksAgent Memory Service stores two types of records:Memories are durable, semantically searchable knowledge entries – facts, decisions, summaries, and preferences that an agent should recall across future sessions. When an agent writes a memory, the service automatically generates a vector embedding using SAP HANA Cloud’s NLP engine. At retrieval time, a vector similarity search surfaces the most relevant memories for the current context, even when the query shares no keywords with the stored content.Messages are raw, chronological conversation logs grouped by session. They give agents an auditable, replayable record of every interaction – essential for compliance, handovers, and multi-turn context within an active session.All data is scoped to an (agentID, invokerID) pair, ensuring that memories and messages are isolated per agent and per user, with no cross-contamination between tenants, agents, or users.The service exposes a standard OData REST API with four key endpoints: No proprietary SDK is required. An optional Python client is available via the SAP Cloud SDK for Python for those who prefer it.What Becomes PossibleAgent Memory Service transforms what an agent can do across three dimensions:Continuity across sessions – A user’s preferences, in-progress tasks, and past decisions survive session boundaries. When a user returns, whether hours or weeks later, the agent picks up where they left off. For example – “Customer prefers consolidated invoices” is automatically recalled in every billing cycle, without the user needing to say it again.Leaner, more precise prompts – Rather than appending entire conversation histories to every prompt, agents retrieve only the top-K semantically relevant memories per turn. This keeps prompts lean, reduces LLM costs, and improves response quality.Improvement over time – As memories accumulate and access timestamps update, agents get sharper without retraining. Frequent dispute patterns are recognized and resolved faster. Recurring questions are anticipated and addressed proactively.Human oversight and control – Users can inspect, edit, or delete their memories directly, providing full transparency. Audit teams can trace how a billing decision or dispute was handled, turn by turn, using the message log.Getting StartedAgent Memory Service is available as a standalone SAP BTP service subscription tied to your SAP HANA Cloud instance (Bring Your Own HANA plan).Provision via SAP BTP Service Marketplace – In your SAP BTP subaccount, add the hana plan for SAP HANA Cloud and the hana-agent-memory plan as entitlements. Provision an SAP HANA Cloud instance or use your already existing one (make sure to enable “nlp”: true in the ‘Advanced Setting’ configuration – this is required for semantic search). Then create an Agent Memory Service instance using the ‘hana-agent-memory’ plan, link it to your HANA Cloud instance, and create a service key. The key gives you the clientid, clientsecret, url, and application_url fields needed to authenticate via OAuth 2.0 client credentials.Subscribe to ‘hana-agent-memory’ service plan on SAP BTP Cockpit Call the API directly or use the SAP Cloud SDK for Python – Once provisioned, you can call the OData REST API from any HTTP client. For Python-based agent development, the SAP Cloud SDK for Python provides an optional client that handles authentication and request handling, with ready-to-use patterns for storing memories, running semantic search, logging conversation turns, and managing retention.For the full step-by-step setup guide, API samples, and best practices, see the SAP Help Portal: Agent Memory Service Guide. The complete API reference and interactive sandbox are available on SAP Business Accelerator Hub.ConclusionThe journey toward agentic SAP HANA Cloud has been about making agents smarter at every layer, from the data they discover and query, to the context they carry with them over time.Agent Memory Service addresses the most fundamental limitation in today’s agent architectures: the absence of persistent, governed, enterprise-grade memory. With it, agents stop being one-shot tools and become evolving partners.For developers, this means:A managed, persistent memory layer that works across SAP and non-SAP agent frameworks (LangChain, LangGraph, SAP BAIP, and custom agents)Semantic retrieval powered by SAP HANA Cloud’s vector engine, no external embedding service requiredNative integration in Joule Studio (planned towards 2027) and the SAP Cloud SDK for PythonBuilt-in governance: configurable retention, data isolation, OAuth 2.0 security, and audit loggingWith Agent Memory Service now generally available, SAP HANA Cloud continues to establish itself as the database for SAP’s AI-native software architecture, not just storing data, but helping agents remember, learn, and improve across every interaction.Note on availability: The service is currently available on a bring-your-own SAP HANA Cloud model, where you provision and manage your own HANA Cloud instance. A fully managed offering, where SAP handles the underlying infrastructure entirely, with no HANA Cloud instance required on your end, is planned towards the end of 2026. Stay tuned for updates. Read More Technology Blog Posts by SAP articles
