memory-compact

name	memory-compact
description	Emergency single-pass memory compression when context window pressure is high mid-task. Activate when approaching token limits during active work. For scheduled end-of-phase cleanup, use memory-refresh instead.
license	MIT
compatibility	Works with any filesystem-based AI coding agent
metadata	{"author":"gaai-framework","version":"1.0","category":"cross","track":"cross-cutting","id":"SKILL-MEMORY-COMPACT-001","updated_at":"2026-03-01T00:00:00.000Z","status":"stable"}
inputs	["contexts/memory/index.md","contexts/memory/**/*"]
outputs	["contexts/memory/summaries/*.summary.md","contexts/memory/archive/**","contexts/memory/index.md (updated)"]

Memory Compact

Purpose / When to Activate

Activate when:

• Context window pressure is high
• Memory has grown across many sessions
• A single targeted compression pass is needed

More focused than memory-refresh — this is a single-pass compression operation.

---

Process

1. Select memory by category or tags. Read contexts/memory/index.md. Prioritize categories by: (a) largest file count first, (b) oldest entries first, (c) categories not referenced by the current task last. Under extreme pressure, compact the single largest category only.

2. Classify entries by durability (R7 gate). Before compacting, classify each entry:

   • Durable (decisions, patterns, project, ops, contacts, domains): only entries with explicit supersession markers (> SUPERSEDED by DEC-XX, > RETRACTED, > OBSOLETE — {reason}) may be archived. All other entries are ACTIVE and MUST NOT be archived regardless of file size. Note: decisions are already individual ADR files (decisions/DEC-{N}.md). For other oversized durable files → domain-split, not archive.
   • Ephemeral (sessions): standard compaction applies — summarize and archive.

3. Extract key decisions, constraints, priorities

4. Generate a single summary file replacing multiple entries. Produce one summary file per compacted category using bullet format: one bullet per decision, constraint, or durable fact. Target ≤20% of the original token count. Use the format:

# {Category} — Compact Summary

> Compacted from {N} files on {YYYY-MM-DD}
> Original token estimate: ~{X} | Summary: ~{Y}

## Key Decisions

- {decision 1}
- {decision 2}

## Active Constraints

- {constraint 1}

## Current State

- {fact 1}

5. Archive detailed originals (ephemeral and superseded only). Archive originals to contexts/memory/archive/{category}-{YYYY-MM-DD}.archive.md. If multiple compactions happen on the same day for the same category, append a sequence number: {category}-{YYYY-MM-DD}-02.archive.md.

6. Update memory index

---

Outputs

• contexts/memory/summaries/{category}.summary.md
• contexts/memory/archive/{category}-{YYYY-MM-DD}.archive.md
• contexts/memory/index.md (updated — mandatory)

---

Quality Checks

• One summary replaces many files
• Context remains precise and small
• No active constraints are lost
• Index reflects current state
• Summary preserves all active decisions and constraints from the originals
• Archive files are never deleted — only moved
• No active durable memory entry archived — only superseded/retracted entries may be archived (R7/R7b)

---

Non-Goals

This skill must NOT:

• Create new project knowledge
• Invent or reinterpret decisions
• Delete (only archive) source files

One summary replaces many files. Context stays precise and small.