Features

1. Hook-Driven Capture

Four hooks drive 100% adoption — not 0%. Corrections, decisions, and preferences are captured at the moment they happen, not when the agent feels like saving them.

Hooks:

• PreToolUse — Before the agent reads any file
• Stop — When the agent finishes a task
• UserPromptSubmit — When you correct the agent
• PreCompact — Before context compaction

2. Path-Scoped Recall

A memory about test utilities surfaces when you open test files, not when you open database migrations. Glob inheritance means src/** memories are available everywhere under src/, but src/checkout/** memories only surface in checkout code.

3. Nudge, Not Dump

~20 tokens per nudge versus ~2,000 tokens dumped into the system prompt. The agent is told context exists and decides relevance. Over hundreds of file reads per day, this saves tens of thousands of tokens.

4. File-Per-Memory, Git Syncs

Every memory is a single JSON file in .aide/memories/<layer>/. No separate sync mechanism. Memories are files. Files are committed. Git syncs them. Delete SQLite and it rebuilds from the JSON files.

5. Structured Memory Layers

Four layers keep memories organized:

• preferences — Your coding style
• technical — Codebase facts
• area_context — File and module notes
• guidelines — Team rules

Recall filters by layer and path.

6. Cross-Tool Portability

Claude Code and Cursor supported out of the box. Same memories, same hooks, same MCP server. Switch tools mid-task and your context follows you.

7. Full-Text Search

BM25-ranked search via FTS5. Find any memory by keyword instantly. Falls back to LIKE-based search when FTS5 is unavailable.

8. No Cloud, No Docker, No API Keys

Everything runs locally. One npm package, SQLite for caching, JSON files for persistence. Your memories never leave your machine unless you commit them to your repo.