ellyseum-claude-total-recall

claude-total-recall

Eidetic memory for Claude Code sessions. Scans all memory files on session start and injects a knowledge map so Claude knows what's in each file without reading them all.

The Problem

Claude Code loads MEMORY.md (the index) but not the 80+ memory files it points to. The index says voice.md exists but doesn't say it contains "Deepgram endpointing config" or "ElevenLabs v3 audio tags." So when you ask about voice latency, Claude doesn't know to read voice.md first - it guesses or greps the codebase.

The Fix

A SessionStart hook extracts ## headers from every memory file and injects them as context. Now Claude sees:

voice: Package Status, Architecture, Phase 5: Colony Voice Integration, Phase 5b: WS Streaming Pipeline...
honeycomb: What It Is, Architecture, Event Schema, Routing Mechanics, Competitive Position...

When you mention "event routing," Claude immediately knows to Read honeycomb.md. No guessing, no grep.

Install

# Clone into your plugins directory
cd ~/projects
git clone https://github.com/your-org/claude-total-recall.git

# Add to Claude Code plugin discovery (or manually add to settings)

Claude Code auto-discovers plugins in ~/projects/claude-*/.

How It Works

1. SessionStart fires (also on resume and compaction)
2. Hook scans $PWD/.context/memory/*.md (brain-surgery path) or auto-memory path
3. Extracts ## headers from each file (topic-level signals)
4. Caches the result (~3.5K tokens for 80 files)
5. Injects as additionalContext - Claude sees the knowledge map immediately

Cost

~3,500 tokens per session. That's 0.35% of a 1M context window. The tradeoff: slightly less context space for dramatically better file selection when memory matters.

Cache

Results are cached at ~/.claude/claude-total-recall/knowledge-map.txt. Cache rebuilds when any memory file is newer than the cache. First run scans all files (~200ms for 80 files). Cached runs read one file (~1ms).

Works With

• brain-surgery: If .context/memory/ exists (symlinked from auto-memory), uses that
• Any project: Falls back to Claude Code's auto-memory directory
• Multiple projects: Cache is per-invocation (uses $PWD to find memory dir)