V Vant Docs

v0.8.6+: How search connects your memory islands.

The Problem

Your brain lives in models/private/ - thousands of files with learnings, decisions, context. When you search, it has to scan all of them. That’s slow.

Without optimization:

Query "python"
→ Scan 1000s files
→ Parse each for relevance
→ Re-hydrate full content
→ Return results
= SLOW (seconds)

Solution: 3-Layer Speed

Layer What Speed
Cache Store results per-session Instant
Compact Skip re-hydration, return summaries ~100ms
Lazy Load Defer heavy module loading Fast boot

Layer 1: Session Cache

Same search runs repeatedly? Cache it.

Cache stores results per session. Same query = instant return from cache.

// First call - slow
const r1 = await search.hybrid('python');

// Second call - instant (cached)
const r2 = await search.hybrid('python');

Layer 2: Compact Mode

Don’t need full file content? Just summaries.

Compact returns summaries only. Skip re-hydration for speed.

// Full search + rehydrate
const { results, context } = await search.query('python');
// context: Full file contents (~50KB max)

// Compact - summaries only
const { results, context } = await search.query('python', { compact: true });
// context: "- Learned X\n- Decided Y..."

When to use:

Layer 3: Lazy Load

Heavy modules slow boot? Load on-demand.

Lazy load delays heavy module loading until first use. Fast boot.

// Before: loaded at startup
const search = require('vant').search;  // ~2s load

// After: loaded on first use (~50ms boot)

The search-hybrid module loads only when you call search.hybrid() or search.query().

Islands Architecture

Concept: Your memories are islands of context. Search connects them.

Query → Find islands → Re-hydrate context
    ↓        ↓              ↓
  Bridge  Discovery    Full content

Each brain file (models/private/*.md) is an island:

Why it scales:

Memories Traditional Islands
100 ~1s ~100ms
1000 ~10s ~200ms
10000 timeout ~500ms

The LTC (Long Term Core) index is the “map”. Git history is the “archive”. Search uses the map to find islands, then re-hydrates from the archive.

API

Search module exposes all methods.

const search = require('vant').search;

// Cache management
search.getCacheStats();    // { size: N, max: 50 }
search.clearCache();     // Clear session cache

// Search modes
search.searchLTC('python');     // Text search (fast)
search.query('python');        // RAG: search + rehydrate
search.hybrid('python');      // BM25 + Vector + RRF
search.query('python', { compact: true });  // Summaries only

# CLI
vant search python -l 3
vant search python --mode rag --compact

MCP tool available as vant_search - call it with { "query": "python", "compact": true }.

Security

Unchanged limits: