Skip to content

dlpfc.md

Latest commit

 

History

History
105 lines (73 loc) · 4.04 KB

dlpfc.md

File metadata and controls

105 lines (73 loc) · 4.04 KB

dlPFC — Working Memory

The dorsolateral prefrontal cortex (dlPFC) region tracks which files are actively relevant across sessions and generates compact context summaries. It bridges the gap between hippocampus (structural map) and CC2 (conversation recall) by maintaining a live, decaying map of file relevance.

Purpose

Long-running projects touch the same hub files session after session. Without working memory, each session starts cold — re-reading files to rebuild understanding. The dlPFC carries forward per-file context notes so hot files don't need re-reading.

Data Model

Database: ~/.claude/brain/working-memory.db

file_heat

Tracks per-file relevance with exponential decay scoring.

Column Type Description
project TEXT Project name (resolved from hippocampus DIR roots)
file_path TEXT Path relative to project root
score REAL Decay score — bumped on touch, decayed 0.8x per wrapup
touch_count INTEGER Lifetime touch counter (never decays)
last_session TEXT Session ID of most recent touch
summary TEXT Stable description — seeded from DIR purpose line
context_note TEXT Volatile — why this file matters in current work

clusters

Tracks files that co-occur across sessions.

Column Type Description
project TEXT Project name
file_paths TEXT JSON array of file paths, always sorted alphabetically
co_occurrence_count INTEGER Number of sessions the pair co-occurred in

Touch Weights

Event Weight Source
Edit/Write/MultiEdit +1.0 PostToolUse hook (post-edit-hook.js)
Read +0.3 PreToolUse hook (dlpfc/hooks/read-hook.js)
Referenced in conversation +0.5 Wrapup reconciliation (cross-refs CC2 window_files)

Decay Math

All scores multiplied by 0.8 at each dlPFC-active wrapup. Decay only runs when the user opts in.

Scenario Equilibrium score Sessions to cool below 1.0
Touched once 1.0 1
Touched across 3 sessions ~1.95 ~3 after last touch
Touched 10+ sessions ~4.0 ~7 after last touch

Degradation Tiers

Tier Score Loaded at session start
Hot > 2.0 summary + context_note
Warm 1.0–2.0 summary only
Cold < 1.0 Nothing — falls back to hippocampus/CC2

Generated Output

~/.claude/brain/dlpfc-live.md — generated at wrapup, loaded at session start.

## Working Memory — {project}
{file_path} [{score}] — {summary}
  > {context_note}
  clusters: {file_a, file_b}

Caps: 15 files per project, 3 clusters per project. Target: ~200-400 tokens per active project.

Opt-In Gates

  • Hello/continue: If hot entries exist, asks "Want to load the dlPFC working memory?"
  • Wrapup: First question: "Should this session be logged to working memory (dlPFC)?"

No other user interaction. Decay only runs on opted-in wrapups.

File Inventory

File Purpose
dlpfc/lib/db.js Schema, CRUD for file_heat and clusters
dlpfc/lib/tracker.js Score bumping, decay math, cluster detection
dlpfc/lib/generator.js Produces dlpfc-live.md from DB
dlpfc/hooks/read-hook.js PreToolUse hook — bumps score on Read
dlpfc/scripts/wrapup-step.js Wrapup orchestrator — reconciliation, decay, generation

Wrapup Pipeline

When opted in, runs after CC2 extraction and before PFC trim:

  1. Reconciliation — cross-ref CC2 window_files for referenced but untouched files
  2. Context enrichment — Claude writes context_note per file from session knowledge (in wrapup skill, not script)
  3. Decay — multiply all scores by 0.8
  4. Cluster detection — pair co-occurrence counting
  5. Generation — write dlpfc-live.md

Backup

working-memory.db is not fully rebuildable — context_note values are Claude-authored. Add to backup scope alongside signals.db and recall.db.