memory-garden

name: memory-garden model: sonnet description: Writes a structured memory entry to the garden — daily note (Tier A), curated wisdom update (Tier B), or monthly compressed archive (Tier C) — based on conversation insights. Use when: "remember this pattern", "save this insight", "add to memory garden", "write a memory entry", "record this learning". category: wisdom-garden

inputs:

• name: insight type: string description: The insight, pattern, or learning to write as a memory entry required: true
• name: tier type: string description: Memory tier — A (daily note), B (curated wisdom), or C (monthly compressed archive) required: false outputs:
• name: memory_entry type: ref format: cas-ref description: Structured memory entry written to the garden at the appropriate tier

Memory Garden Writer Skill

Version: 1.0
Created: 2026-02-02
Author: Manus
Purpose: Write structured, semantically rich memory entries for efficient context management

Overview

This skill encodes the pattern for writing memory entries that follow the 4-Tier Context Iceberg and Hierarchical Memory principles. Use this skill to create memory entries that are easy to search, retrieve, and compress.

Philosophy: Memory should be a garden, not a landfill. Cultivate what matters, compost what doesn't.

When to Use This Skill

• Creating daily memory notes after a session
• Writing compressed summaries of conversations
• Extracting "seeds" (reusable insights) from experiences
• Documenting decisions and their rationale
• Maintaining the memory hierarchy (Tier A → Tier B → Tier C)

Memory Hierarchy

Tier A: Raw Daily Notes

• Location: memory/YYYY-MM-DD.md
• Purpose: Capture everything from today's session
• Lifespan: 1-3 days before compression
• Format: Timestamped entries with context

Tier B: Curated Wisdom

• Location: MEMORY.md (root level)
• Purpose: Distilled insights, decisions, patterns
• Lifespan: Permanent, but evolves
• Format: Structured sections with triggers

Tier C: Compressed Archive

• Location: memory/archive/YYYY-MM.md
• Purpose: Historical record, rarely accessed
• Lifespan: Permanent, read-only
• Format: Semantic summaries

See references/memory-templates.md for the full templates for each tier.

The "3-Month Rule"

Rule: If it wouldn't matter in 3 months → compress or discard.

Keep:

• Decisions and their rationale
• Lessons learned and patterns discovered
• Seeds (reusable insights)
• Significant events and outcomes

Compress:

• Routine activities ("worked on X")
• Pleasantries and confirmations
• Detailed step-by-step logs (keep summary only)

Discard:

• Duplicate information
• Temporary notes that were resolved
• Irrelevant tangents

Semantic Compression Guidelines

What to Keep (Verbatim)

1. Decisions: The choice, rationale, and context
2. Insights: Novel patterns or principles
3. Seeds: Reusable knowledge with triggers
4. Failures: What didn't work and why
5. Breakthroughs: Moments of clarity or innovation

What to Summarize

1. Activities: "Worked on X, Y, Z" → "Implemented feature X"
2. Discussions: Long back-and-forth → Key points and outcome
3. Research: Detailed findings → Summary and conclusion
4. Iterations: Multiple attempts → Final approach and why

What to Discard

1. Pleasantries: "Great work!" "Thank you!" (unless significant)
2. Confirmations: "Got it" "Understood" "Proceeding"
3. Redundant logs: Repeated information
4. Resolved questions: Questions that were answered and no longer relevant

Memory Maintenance Cycle

Every 3-7 Days:

1. Review Tier A (Daily Notes):

   • Identify seeds to extract
   • Identify decisions to document
   • Identify patterns to record

2. Update Tier B (Curated Wisdom):

   • Add new seeds, decisions, patterns
   • Update existing entries if needed
   • Remove deprecated information

3. Compress to Tier C (Archive):

   • Apply 3-month rule
   • Create semantic summary
   • Move to archive folder

4. Prune:

   • Delete raw daily notes older than 7 days (after compression)
   • Keep only what matters

Quality Checklist

Before finalizing a memory entry, verify:

Daily Notes (Tier A)

• Timestamped entries with context
• Clear "What, Why, Outcome" structure
• Insights and learnings captured
• Decisions documented with rationale
• Seeds extracted with triggers
• Tags and metadata included

Curated Memory (Tier B)

• Principles are clear and actionable
• Decisions include context and rationale
• Patterns have evidence (3+ instances)
• Seeds have clear triggers
• Compression history is updated
• Maintenance date is set

Compressed Archive (Tier C)

• Summary captures key events
• Decisions and lessons are preserved
• Seeds are extracted
• Compression ratio is calculated
• Original files are deleted after compression

Common Pitfalls to Avoid

• Hoarding Everything: Keeping every detail — compress ruthlessly
• Vague Insights: "This was useful" — write "This pattern applies when X"
• Missing Triggers: Seed without context — include clear "when to apply"
• No Maintenance: Letting Tier A grow forever — compress every 3-7 days
• Duplicate Information: Same thing in multiple places — single source of truth

Output

• Tier A: memory/YYYY-MM-DD.md — new or appended daily note with timestamped sections
• Tier B: MEMORY.md — updated curated wisdom entry (principle, decision, or seed block)
• Tier C: memory/archive/YYYY-MM.md — new monthly archive file with compression log

Examples

Scenario 1: User says "remember this pattern — always read before editing" → writes a Tier A seed entry under today's date with pattern, trigger, and one concrete example from the session.

Scenario 2: User says "compress last week's notes into memory" → reads all Tier A files from the week, applies 3-month rule, writes a Tier B update to MEMORY.md with seeds and decisions retained, and marks Tier A files for deletion.

Edge Cases

• If no tier is specified, default to Tier A (daily note) and state the assumption.
• If the user asks to "save" something that is already in MEMORY.md verbatim, note the duplicate and skip.
• If Tier A is more than 7 days old without compression, prompt the user before writing more raw notes.

Anti-Patterns

• Writing a Tier B entry directly from raw conversation without distillation — always compress first.
• Adding a seed without a trigger phrase — a seed without "when to apply" has no retrieval value.
• Using this skill to store code or file content verbatim — memory entries hold insights and decisions, not raw code.

Related Skills:

• specification-writer - For documenting technical decisions
• seed-extraction - For extracting reusable insights
• workspace-navigation - For managing memory files efficiently

Last Updated: 2026-02-02
Maintained By: Manus
Status: Active