Agentic Memory

Agentic Memory provides persistent recall of past interactions and accumulated knowledge about each individual your agents serve. Data 360 processes agent conversations to extract and store structured knowledge that includes session summaries, profile facts, and preferences. This information builds over time into a comprehensive understanding of each user. Every future interaction builds on previous interactions, regardless of which agent the user interacts with.

REQUIRED EDITIONS

Available in: All editions supported by Data 360. See Data 360 edition availability.

When an agent converses with a user, conversation data flows through Data 360’s real-time pipeline.

1. The agent sends conversation entries (messages, traces) to Data 360 via the conversation APIs.
2. Data 360 persists the conversation data and indexes it for retrieval.
3. At a periodic cadence during the session, Data 360 processes accumulated messages to extract memories.
4. Data 360 stores extracted memories in the AiParticipantInsight DMO and links them to the relevant individual or unified individual.
5. Data 360 makes memories available for context retrieval within seconds of extraction.

All memory is scoped to a data space and linked to a unified individual. A unified individual is a single resolved identity that persists across agents, sessions, and channels.

Data 360 extracts and stores three types of memory from agent conversations.

• Session summaries: Compressed records of previous conversations including topics discussed, decisions made, actions taken, and outcomes reached. Session summaries provide recall of prior interactions without replaying full transcripts.
• Profile facts: Identity facts about the user distilled from conversations, such as their name, role, company, relationships, key dates, deadlines, and other factual information. Profile facts persist across sessions and build a growing factual record of each individual.
• Preferences: Stated and inferred user preferences across any subject, such as communication style, product choices, scheduling habits, and tool preferences. Preferences are scored by degree of like or dislike and improve personalization over time.

Data 360 extracts memories from agent conversations using an LLM-powered extraction pipeline. Extraction runs at a periodic cadence during the session at intervals determined by message volume and elapsed time. Memory extraction is language-agnostic. Data 360 extracts and stores memories in the language used during the session. For example, a conversation in Spanish produces Spanish-language memories.

Note: Memory extraction consumes flex credits for LLM processing. Each extraction cycle processes the accumulated messages since the last cycle and produces one or more memory entries.

Agentic Memory respects Data 360’s standard governance model.

• User-level access control: Object-level security (OLS), field-level security (FLS), and record-level security (RLS) apply to all memory operations including conversation retrieval, context assembly, and memory extraction.
• Agent-level access control: Determines which agents can read which memories. Each agent reads its own data only.
• Identity scoping: Data 360 links all memory to the unified individual. This ensures consistent visibility across agents. The profile DMO selection in a data space affects all agents in that data space.

• Salesforce Help: Data Governance in Data 360

j

Agentic Memory and Context in Data 360

Agentic Memory and Context Setup

Agentic Memory

Conversation Management

Summary Generation

Agentic Context