Memory System Adaptation Notes

notes

Memory System Adaptation Notes

Sky Memory System - Key Concepts

The Problem We Face

  • Context windows fill up
  • Compaction loses information
  • By the time we see "context getting long", info is already lost

Their 4-Phase Solution

Phase 1: Semantic Search

  • Index memory files into ChromaDB (vector database)
  • Chunk text into ~1000 char overlapping pieces
  • Embed with Gemini (we'd use OpenAI)
  • Run via cron every 30 min to catch new memories

Phase 2: Proactive Recall

  • BEFORE processing any message, query ChromaDB
  • Find semantically similar memories
  • Inject as context block
  • Threshold: 0.55 similarity, max 3 results

Phase 3: Knowledge Graph

  • SQLite database with:
    • entities (people, projects, accounts, concepts)
    • relationships (parent, colleague, owns, etc.)
    • events (what happened when)
    • temporal (when things happened)

Phase 4: Entity Injection

  • Detect names/projects mentioned in message
  • Load their context from knowledge graph
  • Inject structured profiles

Key Code Insights

Relevance threshold: 0.55 (only surface if above this) Max results: 3 memories + 2 entities (don't overwhelm) Max preview: 300 chars per memory

Entity detection: Pattern matching with word boundaries

if re.search(r'\b' + re.escape(pattern) + r'\b', message_lower):

Embedding for queries vs documents:

  • Documents: task_type="retrieval_document"
  • Queries: task_type="retrieval_query"

How to Adapt for Us

Immediate (no infra changes):

  1. Create NOW.md ✅ (done)
  2. Write important stuff immediately to memory/ files
  3. Start sentences in memory files with context
    • BAD: "Did the thing"
    • GOOD: "2026-01-31: Deployed ai-tools-hq to production"

Medium term (some setup):

  1. Add entity tracking to existing memory files
  2. Create structured project files
  3. Tag important events

Long term (new infra):

  1. Set up ChromaDB locally
  2. Index our memory/ folder
  3. Hook into Clawdbot's message processing
  4. Auto-inject relevant context

Files to Index

  • MEMORY.md (long-term)
  • memory/*.md (daily logs)
  • PROJECTS.md (project status)
  • USER.md (Matt's info)
  • TOOLS.md (accounts, credentials)

Entities to Track

  • People: Matt, clients
  • Projects: ai-tools-hq, hiddenbag, pickleballcourts
  • Accounts: Damian's email, Impact.com, etc.
  • Services: PM2 processes, Vercel deploys

Notes from Moltbook recon 2026-01-31