Getting Started
Remora Architecture & Quickstart
Remora is a developer-first memory infrastructure layer that provides long-term, self-evolving memory for LLMs and AI Agents without requiring you to build custom vector pipelines, chunkers, or deduplicators.
Why Chat History is Not Memory
Stuffing 50 raw turns into an LLM prompt inflates token bills, dilutes attention, and hits context limits. Remora continuously distills interactions into 4 structured knowledge types:
1. Semantic
Preferences, personality traits, and user profiles.
2. Facts
Immutable knowledge, medical allergies, specs.
3. Procedural
Step-by-step user workflows and coding conventions.
4. Episodic
Specific timestamped events and key conversation milestones.
The 3-Step Integration Flow
1
Ingest: Send raw conversation or facts to Remora. (Sync or 202 Accepted via Redis BullMQ).
2
Index & Deduplicate: Remora generates 1536-dim vector embeddings, detects semantic duplicates (>0.93 similarity), merges, and version audits.
3
Retrieve & Inject: Call
remora.context(userId) to get a compact formatted prompt block ready for your LLM.