At Dreamforce 2026, Salesforce put a sharp spotlight on one of the biggest unresolved challenges in enterprise AI: while agents can answer, automate and orchestrate, most still lack continuity across interactions.

The continuity gap in enterprise AI

Over the past few years, organizations have invested heavily in AI agents, copilots, automation, orchestration frameworks, prompts and workflows. These systems have become increasingly capable of completing tasks and interacting with enterprise systems. Yet across channels, sessions and handoffs, continuity often breaks down. Customers repeat themselves, agents lose relevant history and enterprises miss the opportunity to turn prior interactions into knowledge that can improve what happens next.

This rwas the central tension Salesforce explored in its session, “Agents That Remember: Memory & Context for Every Interaction.” The key takeaway was that the next leap in agent performance will not come from adding another agent, but from giving agents the foundation to learn, retain what matters and act with the right context.

The problem: AI agents are smart, but they forget

Salesforce illustrated the problem through a familiar customer journey. A visitor engages on a website, later becomes a known customer, is handed off across agents, returns in a new session and eventually reaches a human service representative.

To the customer, these moments are part of one continuous relationship with the brand. To the underlying systems, however, context is often scattered across disconnected interactions.

The session framed this fragmentation through four challenges that continue to limit today’s AI experiences:

Broken continuity

Customers repeatedly explain the same information as conversations move across channels, sessions and agents.

Limited learning across Interactions

Historical conversations may exist, but raw transcripts are not an effective memory system. Tens or hundreds of previous conversations cannot simply be replayed into an LLM each time a customer returns.

Context deficit

Without access to relevant business context, agents often respond with generic answers or incomplete understanding.

Token and integration costs

Many AI implementations repeatedly retrieve information from multiple systems, increasing complexity, latency and cost.

The result is an AI experience that may feel efficient in the moment but does not become more intelligent as the relationship deepens. Customers repeat themselves, agents miss relevant history and enterprises lose the opportunity to turn prior interactions into reusable knowledge.

Salesforce’s answer: agent memory

To address this challenge, Salesforce introduced Agent Memory.

The idea is simple:

Every interaction contributes to a shared intelligence foundation that can be used by future agents, workflows and decisions.

Instead of treating conversations as isolated events, interactions become persistent, reusable knowledge that can improve the next engagement.

Salesforce described Agent Memory as a shared memory layer that captures information across interactions and allows agents to continue from where previous conversations ended.

Memory is not conversation history

One of the most important distinctions made during the session was the difference between conversation storage and memory.

A conversation transcript tells us everything that happened. Memory determines what should be retained, trusted and made reusable.Salesforce demonstrated how multiple conversations can be distilled into reusable intelligence such as:

What happened

Important events, outcomes and resolutions.

Who they are

Customer profile information and relevant attributes.

Preferences

Communication preferences, behavioral patterns and recurring choices.

Rather than forcing future agents to analyze lengthy historical conversations, these distilled memories become reusable knowledge assets that improve service continuity, decisioning and personalization over time.

The three capabilities behind agent memory

Salesforce structured the announcement around three core capabilities.

1. Capture every interaction into shared memory

Every customer interaction becomes an input into a shared memory layer. As customers engage across journeys, the platform continuously builds understanding through:

• Facts
• Preferences
• Learnings
• Summaries
• Profile context

The memory is no longer tied to a single session or a single agent. Instead, every interaction contributes to a growing understanding of the customer and the next interaction benefits from everything learned previously.

2. Activate every agent with full context

The second capability is where the session became particularly interesting.

Salesforce repeatedly emphasized that memory alone is insufficient. Agents also need context. To address this, Salesforce introduced a context assembly approach that brings together information from multiple sources, including:

• Memory
• Data Cloud
• CRM records
• Real-time customer profile information
• Business data
• Web engagement signals

Unstructured knowledge and retrieval layers

The objective is not to deliver every available piece of information but to bring in only the information required for the current interaction.

Salesforce described this through a context assembly process that determines:

• What information is relevant
• Which information the agent is authorized to access
• What context is necessary to answer or act

This allows agents to operate with greater precision while reducing unnecessary token consumption, duplication and latency.

3. Govern everything from day one

The third capability focused on governance.

As memory becomes shared, permissions become increasingly important. Salesforce repeatedly reinforced that memory must remain governed. Not every agent needs to access every piece of information and not every interaction requires complete visibility into a customer’s profile.

The platform’s governance model ensures that:

• Memory remains permission aware
• Data policies follow the memory
• Enterprise trust controls remain enforced
• Sensitive information remains protected

A key message throughout the session was that governance is not an afterthought, but it is built directly into the architecture.

Beyond memory: A foundation for reusable intelligence

What makes this significant is that Salesforce is not positioning Agent Memory as a standalone feature. The session framed memory as part of a broader context foundation, where customer history, enterprise data, knowledge, policies and permissions come together before an agent acts.

That direction aligns with Persistent’s view of the enterprise AI context layer: intelligence becomes useful when it can draw on governed context, institutional knowledge and memory across workflows. Salesforce’s emphasis on memory that persists across interactions maps closely to the enterprise need for reusable knowledge. What carries forward becomes valuable when it can be governed, permissioned and applied in the right business context.

Salesforce’s vision suggests a future where enterprise AI systems are built on three foundations:

• Memory that persists across interactions and captures what matters from prior engagement
• Context that turns enterprise data and knowledge into intelligence agents can reuse
• Governance that ensures this intelligence is accessed securely, appropriately and in the right business context

What this signals for enterprise AI

The concepts showcased at Dreamforce reinforce a broader direction in enterprise AI: agents will need more than task-level autonomy. They will need a trusted context foundation that brings together enterprise knowledge, memory, data and governance into a reusable layer for AI.

This is where Salesforce’s session connects to the larger industry shift toward agent ecosystems. As Persistent has outlined in its view of the enterprise AI context layer, the differentiator will not be the agent alone, but the governed intelligence that surrounds it.

Salesforce’s emphasis on minimum, relevant and permission-aware context for every interaction reflects the same movement toward AI systems that can carry knowledge forward without compromising trust.

As AI moves from individual agents to connected agent ecosystems, the real differentiator will be the intelligence layer behind them: governed memory, reusable knowledge and contextual understanding that allow every interaction to make the next one better.