Eidolon Search

FTS5-based memory search for AI Agents. Index markdown files once, search with 90%+ token savings.

Quick Start

1. Index memory files (once)

CODEBLOCK0

Example:
CODEBLOCK1

This creates a SQLite database with FTS5 full-text index of all .md files in the directory (recursive).

2. Search

CODEBLOCK2

Example:
CODEBLOCK3

Default limit: 10. Default db_path: INLINECODE1

Output: matching snippets with file paths and relevance scores.

3. Re-index when files change

Run build-index.py again. It rebuilds the index from scratch (fast, <1 second for typical workspaces).

When to Use

• - Memory search: Find specific information across many daily notes or memory files
• Token-limited contexts: When reading all files would exceed context limits
• Repeated searches: Index once, search many times
• Large workspaces: 10+ markdown files with cumulative size >50KB

When NOT to Use

• - Single small file (<5KB): just read it directly
• Need semantic/meaning-based search: FTS5 is keyword-based only
• See references/QUALITY.md for known limitations

Search Tips for Agents

FTS5 is keyword-based. Improve results by:

• - Use specific terms: "Jetson Orin" not "hardware plans"
• Use OR for synonyms: "car OR vehicle OR automobile"
• Use quotes for phrases: INLINECODE3
• Try multiple queries if first attempt returns nothing
• Check file paths in results to read full context when needed

Benchmarks

• - Token savings: 90%+ (measured 93-98.9%)
• Speed: 15x faster (measured 10-20x)
• Details: references/PERFORMANCE.md

Run benchmarks yourself:
CODEBLOCK4

DB Schema

CODEBLOCK5

Direct SQL access:
CODEBLOCK6