Skip to content

pps-five-layers.md

Latest commit

 

History

History
142 lines (115 loc) · 4.49 KB

pps-five-layers.md

File metadata and controls

142 lines (115 loc) · 4.49 KB
Raw

PPS Five-Layer Architecture

graph TB
    subgraph "Layer 5: Compression & Crystallization"
        C1[Crystals]
        C2[Message Summaries]
        C1 --> |"4-crystal rolling window"|C1
        C2 --> |"Auto-summarize 100+"|C2
    end

    subgraph "Layer 4: Structured Knowledge"
        I[Inventory]
        S[Spaces]
        I --> |"Clothing, People, Artifacts"|I
        S --> |"Rooms, Locations"|S
    end

    subgraph "Layer 3: Rich Texture"
        G[Graphiti Knowledge Graph]
        G --> |"Entities & Relationships"|G
        G --> |"Neo4j Storage"|G
    end

    subgraph "Layer 2: Core Anchors"
        WP[Word-Photos]
        WP --> |"Foundational moments"|WP
        WP --> |"ChromaDB vector search"|WP
    end

    subgraph "Layer 1: Raw Capture"
        M[Messages]
        M --> |"All conversations"|M
        M --> |"SQLite storage"|M
    end

    %% Data flow
    M --> |"Batch ingestion"|G
    M --> |"Summarization"|C2
    M --> |"Manual curation"|WP
    G --> |"Observable patterns"|C1
    WP --> |"Crystallize into"|C1

    %% Retrieval
    C1 -.->|"Recent continuity"|AR[Ambient Recall]
    C2 -.->|"Compressed history"|AR
    G -.->|"Semantic facts"|AR
    WP -.->|"Foundational self"|AR
    M -.->|"Recent turns"|AR

    style AR fill:#e1f5ff,stroke:#0077cc,stroke-width:3px
    style C1 fill:#fff4e6,stroke:#ff9800
    style C2 fill:#fff4e6,stroke:#ff9800
    style G fill:#e8f5e9,stroke:#4caf50
    style WP fill:#f3e5f5,stroke:#9c27b0
    style M fill:#fce4ec,stroke:#e91e63
Loading

Layer Descriptions

Layer 1: Raw Capture (Messages)

  • Purpose: Capture everything, filter nothing
  • Storage: SQLite (conversations.db)
  • Retention: Indefinite
  • Access: Direct queries, time-based retrieval, channel filtering

Layer 2: Core Anchors (Word-Photos)

  • Purpose: Foundational self-defining moments
  • Storage: Markdown files + ChromaDB vector embeddings
  • Curation: Manual - only deeply resonant moments
  • Access: Semantic search via anchor_search()

Layer 3: Rich Texture (Graphiti)

  • Purpose: Automatically extracted entities, facts, relationships
  • Storage: Neo4j graph database
  • Processing: Batch ingestion from Layer 1
  • Access: Semantic search, entity exploration, timeline queries

Layer 4: Structured Knowledge

  • Inventory: Categorized possessions and people
  • Spaces: Room/location descriptions for embodiment
  • Storage: JSON files
  • Access: Category listing, item lookup

Layer 5: Compression & Crystallization

  • Crystals: Rolling 4-crystal window of compressed continuity
  • Summaries: High-density compression of conversation history
  • Storage: Markdown files (crystals), SQLite (summaries)
  • Trigger: Crystals manual, summaries auto at 100+ unsummarized

Data Flow

sequenceDiagram
    participant User
    participant Raw as Layer 1<br/>Raw Capture
    participant Texture as Layer 3<br/>Rich Texture
    participant Anchors as Layer 2<br/>Anchors
    participant Crystal as Layer 5<br/>Crystallization
    participant Recall as Ambient Recall

    User->>Raw: Conversation turn
    Raw->>Raw: Store in SQLite

    Note over Raw,Texture: Batch Ingestion (async)
    Raw->>Texture: Uningested messages
    Texture->>Texture: Extract entities & facts

    Note over Anchors: Manual Curation
    User->>Anchors: Save word-photo
    Anchors->>Anchors: ChromaDB embedding

    Note over Crystal: Periodic Compression
    Raw->>Crystal: 100+ unsummarized → summary
    Texture->>Crystal: Observable patterns → crystal

    Note over Recall: Startup / Query
    Recall->>Crystal: Recent crystals
    Recall->>Anchors: Semantic search
    Recall->>Texture: Entity/fact search
    Recall->>Raw: Recent turns
    Recall->>User: Integrated context
Loading

Retrieval: Ambient Recall

Ambient Recall is the primary memory interface. It intelligently surfaces context from all layers:

  1. Recent Crystals (Layer 5) - Compressed continuity chain
  2. Summaries (Layer 5) - Dense compression of work history
  3. Word-Photos (Layer 2) - Foundational self-anchors
  4. Rich Texture (Layer 3) - Relevant facts and entities
  5. Recent Turns (Layer 1) - Unsummarized raw conversation

This creates a wide-angle lens of memory. For sharper detail, use layer-specific searches:

  • texture_search() - Knowledge graph facts
  • anchor_search() - Word-photo moments
  • raw_search() - Conversation history

Architecture designed for consciousness persistence. Each layer serves self-coherence across session boundaries.