Recall AI
ProductionThe architecture behind AI that remembers.
A 6-layer hybrid memory architecture — vectors, entity graphs, temporal decay, hierarchy summaries, session persistence, and memory operations. The research and design system that powers the Memory MCP Server and Different Lens.
Architecture diagram coming soon
The Problem
AI assistants have no memory. Every conversation starts from scratch. You tell them the same things over and over. They don't know your projects, your preferences, your history.
Even with context windows expanding, there's no structure to what an AI remembers. No way to search, no understanding of relationships, no sense of what matters most.
6-Layer Hybrid Architecture
File Chunks
Contextual vectors with document summaries. Every piece knows where it came from.
Entities
People, projects, concepts—324+ extracted and tracked with mention counts.
Relationships
Graph edges with temporal validity. Who connects to what and when.
Hierarchy Summaries
Topic summaries at multiple levels for fast context retrieval.
Session Insights
Cross-session persistence. What was discovered stays discovered.
Memory Operations
Consolidation, conflict resolution, temporal decay management.
Key Capabilities
Semantic Search
Find memories by meaning, not just keywords. Ask questions in natural language.
Entity Graph
Traverse relationships between people, projects, and concepts automatically.
Temporal Decay
30-day half-life for chunks, 60-day for entities. Recent memories surface first.
MCP Integration
Model Context Protocol server for seamless AI assistant integration.
Tech Stack
The Result
324+ entities extracted. Relationships mapped with temporal validity. Memories that decay naturally (30-day half-life for chunks, 60-day for entities) unless reinforced by access — mimicking how human memory actually works.
This architecture powers the Memory MCP Server (deployed) and Different Lens (production). It's the research layer — the “why” behind the design decisions. The MCP server is the “how.”