Memory Gardener

Automatic consolidation of ephemeral memory into long-term storage.

Problem

• Daily memory files (memory/YYYY-MM-DD.md) accumulate
• Important insights get buried in noise
• Manual consolidation to MEMORY.md is tedious
• MCP memory (the vector store) exists but isn't systematically seeded

Vision

A cron job or heartbeat task that:

1. Reads yesterday's daily file (or files from last N days)
2. Extracts significant events using lightweight classification:
   • Decisions made
   • Lessons learned
   • User preferences discovered
   • Infrastructure changes
   • Project milestones
3. Deduplicates against existing MEMORY.md entries
4. Proposes additions (or auto-appends in low-risk mode)
5. Stores to MCP memory with appropriate tags for semantic search

Implementation Sketch

# memory_gardener.py

import datetime
from pathlib import Path

def extract_significant(daily_content: str) -> list[dict]:
    """
    Parse daily file, return list of {category, content, importance}

    Categories: decision, lesson, preference, milestone, infrastructure
    Importance: 1-5 (5 = definitely keep)
    """
    # Could use local LLM (qwen2.5-7b) for classification
    # Or simple regex/keyword matching for v0
    pass

def dedupe_against_memory(items: list[dict], memory_content: str) -> list[dict]:
    """Remove items already captured in MEMORY.md"""
    # Fuzzy string matching or embedding similarity
    pass

def append_to_memory(items: list[dict], memory_path: Path):
    """Add new items to appropriate sections of MEMORY.md"""
    pass

def seed_mcp_memory(items: list[dict]):
    """Store to the vector store via the tool proxy"""
    # the tool proxy call memory.store_memory content="..." tags="..."
    pass

Trigger Options

1. Cron job — Daily at 3 AM (sessionTarget=isolated, agentTurn)
2. Heartbeat task — "If yesterday's file exists and not yet processed..."
3. Manual — /garden command

MVP

Start simple:

• Run weekly during heartbeat
• Just flag files older than 7 days that haven't been reviewed
• Human decides what to keep

Then iterate toward automation.

Open Questions

• How aggressive should auto-consolidation be?
• What's the forgetting curve? (When is it OK to lose daily details?)
• Should MCP memory be source of truth, or mirror of MEMORY.md?

Related Dreams

• Cognitive Architecture — Memory Layer automation in the unified system
• Context Archaeology — Memory reconstruction from different angles
• Working Memory Protocol — How consolidated memories flow back into context
• Machines That Think — Minsky on self-images evolving over time

Status: Dream. Not implemented.