F041: Memory Compaction

Status: Proposed Priority: High Inspired by: Beads (steveyegge/beads) - semantic “memory decay” that summarizes old closed tasks

Problem

As decision count grows (currently 182+), loading full decision history into agent context becomes expensive and noisy. Agents waste context window on resolved decisions that could be summarized. There’s no mechanism to distinguish between “active knowledge” and “historical record.”

Solution

Implement semantic compaction that progressively summarizes old, resolved decisions while preserving their calibration value and key learnings.

Compaction Levels

Level	Age	Content
Full	< 7 days	Complete decision with all reasoning, traces, context
Summary	7-30 days	Decision text, outcome, key pattern, confidence vs actual
Digest	30-90 days	One-line summary grouped by category
Wisdom	90+ days	Statistical aggregates + extracted principles

Compaction Process

1. Trigger: Scheduled (daily) or on-demand via API
2. Summarize: LLM-generated summary preserving decision essence
3. Preserve: Raw data always kept in storage; compaction only affects query responses
4. Protect: Decisions with preserve: true or unreviewed outcomes skip compaction

API

cstp.compact          - Run compaction cycle
cstp.getCompacted     - Get decisions at appropriate compaction level
cstp.setPreserve      - Mark decision as never-compact
cstp.getWisdom        - Get category-level distilled principles

Example Output

Wisdom level (90+ days, architecture category):

{
  "category": "architecture",
  "decisions": 45,
  "success_rate": 0.93,
  "key_principles": [
    "Manual type resolution beats annotation magic (3 confirmations)",
    "Search-first prevents duplicate work (8 confirmations)",
    "Parallel independent reasons > single strong argument (5 confirmations)"
  ],
  "common_failure_mode": "Skipping pre-decision query (4 failures)"
}

Phases

1. P1: Time-based compaction levels in query responses
2. P2: LLM-generated summaries for summary/digest levels
3. P3: Wisdom extraction - cross-decision principle mining
4. P4: Configurable compaction policies per agent/category

Integration Points

• F002 (Query): Compaction level affects query response size
• F009 (Calibration): Compacted decisions retain calibration data
• F024 (Bridge Definitions): Bridge summaries survive compaction
• F034 (Decomposed Confidence): Confidence components inform compaction priority