Internal Knowledge Base

Hindsight Memory Engine • TEMPR Retrieval • CARA Reasoning

Automated Ingestion Pipeline
📹
Zoom Recording
Cloud Storage Upload
→
🎙️
Transcription
Speaker Diarization
→
🧬
Fact Extraction
TEMPR Pipeline
→
✨
AI Enrichment
Summary + Actions
Hindsight Four-Network Memory Architecture 91.4% LongMemEval
🌍
World Network
Objective facts: decisions, data points, specifications. Third-person truths independent of any agent.
📝
Experience Network
First-person records of AI actions: questions answered, analyses performed, recommendations given.
🔭
Observation Network
Auto-consolidated knowledge: patterns and trends synthesized from raw facts with evidence tracking.
💭
Opinion Network
Subjective beliefs with confidence scores and timestamps. Updates as new evidence arrives.
TEMPR Multi-Dimensional Retrieval
Example Query
"What did our team decide about the pricing model in Q3, and has anyone revised that since?"
🧠
Semantic Search
Vector cosine similarity for conceptually related content
🔤
Keyword Search
BM25 exact matching for names, terms, identifiers
🕸️
Graph Traversal
Multi-hop entity walks across knowledge graph
⏰
Temporal Search
Date parsing, range filtering, temporal links
🎭 CARA Reasoning
Mission: Natural language identity defining priorities
Directives: Hard rules (e.g., "Always cite source")
Disposition: Tunable traits (skepticism, empathy)
Priority: Mental Models → Observations → Raw Facts
Access Interfaces
💬
Chat Interface
Natural language queries with source citations
🔌
MCP Integration
Claude Code / Desktop access via Model Context Protocol
📚
Meeting Archive
Browse, search, quick-summary popups
📋 Component Details
Technology Stack
Hindsight (Vectorize.io) Zep Graph PostgreSQL + pgvector Whisper / Deepgram Claude (Anthropic) MCP Protocol Cross-Encoder Reranking Reciprocal Rank Fusion