Memory Extraction
Data 360 extracts session summaries from your agent conversations through a real-time memory extraction pipeline. Memory extraction processes conversations at regular intervals, capturing structured knowledge that persists across sessions and becomes immediately available for context retrieval in future conversations.
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.
How Memory Extraction Works
Data 360 extracts memories from conversation data periodically during the session. Extraction runs at intervals determined by message volume and elapsed time, balancing freshness with efficiency.
1PAST(memories) + PAST_OVERLAP(messages) + PAST_SUMMARY(conversations) + NEW(messages) + NEW(traces) = NEW(memories)Each extraction cycle considers both new messages and previously extracted memories. This enables the pipeline to:
- Consolidate overlapping information across cycles.
- Remove duplicate memories.
- Resolve conflicts when newer information contradicts older information (recency wins).
- Build progressively richer and more accurate memories over time.
Extraction Triggers
| Trigger | Description |
|---|---|
| Size-based | Triggered when accumulated unprocessed messages reach approximately 50% of the LLM context window. This is the primary trigger during active conversations. |
| Time-based | Triggered after a configurable duration since the last extraction. Ensures memories are generated from low-volume conversations. |
| Session close | A final extraction cycle runs when a session is closed, processing any remaining unprocessed messages. |
What Gets Extracted
Each extraction cycle can produce these memory types.
| Memory Type | Scope | What It Captures |
|---|---|---|
| Conversation Summary (Episodic) | Conversation | Topics discussed, decisions made, actions taken, outcomes reached, key turning points. |
| Profile Fact (Semantic) | Individual | Name, role, company, relationships, key dates, technical environment, business context. |
| Profile Preference (Semantic) | Individual | Communication style, product choices, scheduling habits, tool preferences. |
Profile facts persist indefinitely. Each preference includes a confidence score (0.0–1.0) and decays over time unless reinforced, with explicitly stated preferences receiving higher confidence than inferred ones.
Considerations
- LLM processing: Memory extraction involves calling an LLM with the extraction prompt that generates structured output containing summaries. This consumes LLM usage credits.
- Memory storage: Data 360 stores all extracted memories as conversation memories.
- Language support: Memory extraction is language-agnostic. The LLM processes messages in whatever language the conversation uses and produces memories in that same language. You don’t need a separate translation step.
j