Memory

Memory

A hybrid search memory system for your project. Save important information and find it later using semantic search.

First run: If slow on first use, run bunx memory-search --warmup to pre-download the embedding model (~300MB).

Memory File Structure

project/
├── MEMORY.md           # Long-term: preferences, patterns, decisions
└── memory/
    └── YYYY-MM-DD.md   # Daily: session notes, context, progress

Searching Memory

When the user asks about past decisions, preferences, or wants to recall information:

bunx memory-search "QUERY_HERE"

Search Examples

bunx memory-search "database choice decision"
bunx memory-search "error handling pattern"
bunx memory-search "user preferences coding style"
bunx memory-search "authentication implementation"

Output Format

Returns matching snippets with:

• path: File where the match was found
• lines: Line range (startLine-endLine)
• score: Relevance score (0-1)
• snippet: The matching text

After Searching

1. Present the relevant results to the user
2. If they want more detail, use Read to get the full file content
3. If no results found, offer to search with different terms

Saving to Memory

When the user says "remember this", "save this", "note that", or wants to store information:

Decide Which File

MEMORY.md (Permanent):

• User preferences ("I prefer TypeScript")
• Project decisions ("We chose PostgreSQL for X reason")
• Coding patterns ("Always use async/await")
• Architecture decisions
• Important URLs, contacts, credentials references

memory/YYYY-MM-DD.md (Daily):

• What was worked on today
• Bugs found and fixed
• Ideas to explore later
• Meeting notes
• Temporary context

How to Save

1. Read existing file first (if it exists) to avoid overwriting
2. Append new content with a timestamp or section header
3. Use clear, searchable language (will be vector-searched later)
4. Run sync after saving to update the search index

bunx memory-search --sync

Example: MEMORY.md

## User Preferences

- Prefers Bun over Node for TypeScript projects
- Uses pnpm as package manager
- Likes minimal dependencies

## Project Decisions

### 2024-01-15: Database Choice
Chose SQLite over PostgreSQL because:
- Single-user application
- No need for concurrent writes
- Simpler deployment

Example: memory/2024-01-15.md

# 2024-01-15

## Session Notes

### 10:30 - Authentication Setup
- Implemented JWT auth flow
- Added refresh token rotation
- TODO: Add rate limiting

### 14:00 - Bug Fix
- Fixed race condition in user creation
- Root cause: missing transaction wrapper

Tips

• Use descriptive queries, not single keywords
• The search is semantic (understands meaning, not just exact words)
• If unsure which file to use, ask the user
• Always sync after saving new content