Memory System

Three-layer architecture for an agent that wakes up fresh every session

flowchart TD
    CONV["Conversations\nTelegram main session"]
    HB["Heartbeats\nPeriodic background work"]

    subgraph L2["Layer 2 — Daily Notes"]
        DN["memory/YYYY-MM-DD.md\nRaw timeline of events"]
    end

    subgraph L1["Layer 1 — Knowledge Graph"]
        direction TB
        AREAS["life/areas/"]
        PPL["people/"]
        CMP["companies/"]
        PRJ["projects/"]
        ITEMS["items.json\nAtomic timestamped facts"]
        SUMM["summary.md\nLiving snapshot"]
        AREAS --- PPL
        AREAS --- CMP
        AREAS --- PRJ
        PPL --> ITEMS
        PPL --> SUMM
        ITEMS -->|"weekly\nsynthesis"| SUMM
    end

    subgraph L3["Layer 3 — Tacit Knowledge"]
        MEM["MEMORY.md\nPatterns, preferences,\nlessons learned"]
    end

    CONV -->|"writes events"| DN
    HB -->|"extracts\ndurable facts"| ITEMS
    DN -.->|"facts flow\nupward"| ITEMS
    DN -.->|"periodic\nreview"| MEM

    SUMM -.->|"context\nat session start"| CONV
    MEM -.->|"loaded\nat session start"| CONV
    DN -.->|"recent days\nloaded"| CONV

The Three Layers

Layer 1: Knowledge Graph
life/areas/{type}/{entity}/

Entity-based storage that compounds forever. Each person, company, or project gets a folder containing:
— summary.md: living snapshot, rewritten weekly
— items.json: atomic timestamped facts

Never delete — supersede instead.
Layer 2: Daily Notes
memory/YYYY-MM-DD.md

Raw timeline — what happened, when. Written continuously during conversations and heartbeats. Durable facts are later extracted upward into Layer 1. The journal.
Layer 3: Tacit Knowledge
MEMORY.md

Patterns, preferences, lessons learned — facts about Echo, not the world. Curated wisdom distilled from daily notes. The index that points to topic-specific files.

The Compounding Engine

The flywheel: Conversation → Facts extracted → Graph grows → Weekly synthesis → Better context → Better responses → More conversation

Real-Time Extraction
During heartbeats: check conversations, extract durable facts (relationships, status changes, milestones), write to entity items.json. Skip casual chat and temporary info.
Weekly Synthesis (Sunday)
For each entity with new facts: load summary.md + active facts from items.json, rewrite summary.md for current state, mark contradicted facts as superseded.

Atomic Fact Schema

{
  "id": "entity-001",
  "fact": "The actual fact",
  "category": "relationship | milestone | status | preference",
  "timestamp": "YYYY-MM-DD",
  "source": "conversation",
  "status": "active | superseded",
  "supersededBy": "entity-002"
}
Core Principle: Echo wakes up blank every session. There is no hidden state, no persistent process, no background memory. The only thing that survives between sessions is what's written to disk. If you think it, write it. Text > brain.