memorytrace 0.3.1

AI agent memory system with SQLite+FTS5, MCP integration, and quality gates

Engram

AI agent memory that learns across conversations.

Install

pip install memorytrace

Use

engram save "Minseong Jeong is the Space King of Galaxy Corp"
engram find "Space King"
engram who "Minseong Jeong"

That's it. 3 commands. No config, no API keys, no database setup.

All Commands

engram save "text"                   Save information (auto-extracts entities & facts)
engram find "query"                  Search memory
engram who "name"                    Look up a person/org
engram remember "name" "fact"        Manually remember something
engram all                           List everything
engram status                        Health check
engram forget "name"                 Delete an entity
engram export                        Export to Markdown files

Examples

# Store meeting notes
engram save "Met Alice Johnson, VP of Engineering at Google. Discussed cloud AI partnership."

# Store Korean
engram remember "정민성" "AI와 우주 탐사에 관심이 많다"

# Search later
engram find "Google partnership"
engram who "Alice Johnson"

# Check status
engram status

What It Does Automatically

• Extracts people & organizations from text (English, Korean, Chinese, Japanese)
• Masks PII — phone numbers, credit cards, emails become [REDACTED]
• Deduplicates — same fact won't be stored twice
• Detects conflicts — "CEO" then "CTO" for same person gets flagged
• Generates summaries — auto-summary when you don't provide one
• Learns across sessions — every conversation builds on previous ones

For Claude Code Users

Add slash commands to use memory inside Claude Code:

# Copy command files
mkdir -p ~/.claude/commands
# See docs/04-usage-guide/02-claude-code-setup.md for details

Then use: /memory-save, /memory-find, /memory-who, /memory-remember, /memory-status

For Developers (Python SDK)

from engram.integrations.sdk import EngramSDK

with EngramSDK() as sdk:
    sdk.store("Minseong Jeong is the Space King of Galaxy Corp")
    result = sdk.search("Space King")
    print(result.to_agent_context())

How It Works

Text → NER Extraction → Quality Gate → SQLite+FTS5 → BM25 Search
                              ↓
                     Confidence Scoring
                     PII Masking
                     Duplicate Check
                     Conflict Detection

• Storage: SQLite + FTS5 (zero dependencies, genuine BM25)
• Export: Markdown files for human reading
• Session: Context carries over between conversations

License

MIT