Skip to content

01-overview.md

Latest commit

 

History

History
103 lines (84 loc) · 6.39 KB

01-overview.md

File metadata and controls

103 lines (84 loc) · 6.39 KB

What is SerialMemory?

SerialMemory is a temporal knowledge graph memory system for AI applications. Think of it as a persistent "second brain" that stores, indexes, and retrieves information using semantic search.

Key Features

  • Semantic Search: Find memories by meaning, not just keywords
  • Entity Extraction: Automatically identifies people, organizations, dates, and concepts
  • Relationship Tracking: Builds a knowledge graph of connections between entities
  • Multi-Hop Reasoning: Traverses relationships to find related context
  • Event Sourcing: Full audit trail of all changes with confidence decay
  • Multi-Tenant: Isolated data per tenant with row-level security

Architecture Overview

┌──────────────────────────────────────────────────────────────────┐
│                         Client Applications                       │
│  (Claude Desktop, Custom Apps, SDKs)                             │
└──────────────────────────────┬───────────────────────────────────┘
                               │
                               ▼
┌──────────────────────────────────────────────────────────────────┐
│                         API Layer                                 │
│  ┌─────────────┐  ┌──────────────────┐  ┌──────────────────┐    │
│  │ MCP Server  │  │ REST API         │  │ Dashboard API    │    │
│  │ (Claude)    │  │ (Custom Apps)    │  │ (Self-Service)   │    │
│  └─────────────┘  └──────────────────┘  └──────────────────┘    │
└──────────────────────────────┬───────────────────────────────────┘
                               │
                               ▼
┌──────────────────────────────────────────────────────────────────┐
│                       Core Services                               │
│  ┌─────────────────────────────────────────────────────────────┐ │
│  │  Knowledge Graph Service                                     │ │
│  │  - Memory ingestion with entity extraction                   │ │
│  │  - Semantic search (pgvector)                                │ │
│  │  - Multi-hop graph traversal                                 │ │
│  │  - User persona management                                   │ │
│  └─────────────────────────────────────────────────────────────┘ │
└──────────────────────────────┬───────────────────────────────────┘
                               │
                               ▼
┌──────────────────────────────────────────────────────────────────┐
│                       Data Layer                                  │
│  ┌─────────────────┐  ┌──────────────┐  ┌─────────────────────┐ │
│  │ PostgreSQL      │  │ Ollama       │  │ Redis               │ │
│  │ + pgvector      │  │ (Embeddings) │  │ (Cache/Rate Limit)  │ │
│  └─────────────────┘  └──────────────┘  └─────────────────────┘ │
└──────────────────────────────────────────────────────────────────┘

Core Concepts

Memories

A memory is a piece of information stored in the system. Each memory has:

  • Content: The actual text/information
  • Embedding: A 1536-dimensional vector for semantic search (OpenAI text-embedding-3-small)
  • Entities: Extracted entities (people, organizations, etc.)
  • Confidence: A score that decays over time unless reinforced
  • Layer: Classification (L0_RAW through L4_HEURISTIC)

Entities

Named things extracted from memories:

  • PERSON: People's names
  • ORG: Organizations and companies
  • GPE: Geographic locations
  • DATE: Dates and times
  • SKILL: Technical skills and capabilities
  • PROJECT: Project names

Relationships

Connections between entities:

  • "John" works at "Acme Corp"
  • "Acme Corp" located in "San Francisco"
  • "Project X" uses "Python"

Multi-Hop Search

Starting from a query, the system can:

  1. Find relevant memories
  2. Extract entities from those memories
  3. Find other memories linked to those entities
  4. Repeat for N hops

This enables questions like "Who knows someone who works at Acme?" to be answered by traversing the knowledge graph.

Use Cases

  1. AI Second Brain: Personal knowledge management for AI assistants
  2. Project Context Memory: Per-project isolated memory for development tools
  3. Customer Support: Historical context for support conversations
  4. Research Assistants: Persistent knowledge across research sessions

Getting Started