systems

Systems

The systems lens covers the memory-engine's major subsystems in depth, one concern per page. Start with Architecture for how these fit together end to end, and Glossary for shared vocabulary.

• Retrieval and search — hybrid vector/text search, fusion methods, reranking, and recall profiles.
• Consolidation and novelty — the offline "Dreamer" pipeline that merges, promotes, or invalidates memories, and surprisal-based novelty detection.
• Graph, episodes, and entities — entity/relation extraction, graph traversal, and episode summarization from conversation events.
• Temporal and bitemporal — valid-time vs. transaction-time tracking and point-in-time queries.
• Structured memory and procedures — typed facts and stored playbooks, their stable-handle addressing and active/invalidated/conflicted lifecycle.
• Schema, migrations, and indexes — collection setup, Atlas Search/Vector Search index management, schema validators, capability detection, and data migrations.
• Embeddings and providers — embedding generation, provider integrations, and automated vs. client-side embedding modes.
• Provenance and evidence — tracing search results back to source events and trust/confidence scoring.
• Jobs, telemetry, and sync — the background job queue, change-stream sync, and operational telemetry.

See also packages/memory-engine/index.md for the package's public surface, and security.md for security-specific concerns such as the injection classifier.