This section walks the pipeline from raw conversation to durable, queryable memory: extract, transform, commit — plus the operational concerns that keep writes correct under load.
What you will learn
- The extract → transform → commit pipeline and why each stage exists
- Deduplication, conflict reconciliation, and merging during transforms
- Async processing, durable execution, ordering, and idempotency
- How pipeline output becomes a queryable memory store
Chapters in this section
There are 20 chapters in this part. Open any chapter to read it on its own, or work through them in order.
- 1 The Memory Pipeline: Extract, Transform, Commit
- 2 Extraction: Pulling Facts from Raw Conversation
- 3 Input Shapes: Conversations, Strings, and Pre-Extracted Facts
- 4 LLM-Powered Extraction vs Rule-Based Extraction
- 5 Transformation: Integrating New Information with Existing Memory
- 6 Deduplication During the Transform Stage
- 7 Reconciling Conflicting Memories
- 8 Merging Related Memories Into Higher-Order Knowledge
- 9 Commit Semantics: Why Atomic Writes Matter
- 10 Asynchronous Memory Processing and Why It Matters for Latency
- 11 Durable Execution for Memory Pipelines
- 12 Ordered Processing: Why Sequence Matters in Memory Updates
- 13 Buffering and Batching Raw Memory Input
- 14 Pipeline Steps as a Composable Graph
- 15 Idempotency in Memory Writes
- 16 Run Tracking and Observability of Pipeline Execution
- 17 Designing Pipelines for Different Use Cases
- 18 Continual Learning Pipelines: Memory That Improves Agent Behavior
- 19 LLM-as-Judge Patterns in Memory Pipelines
- 20 From Pipeline Output to Queryable Memory Store