Engram

Give your AI a perfect memory.

Engram remembers everything you tell it—names, relationships, preferences, conversations—and recalls exactly what's relevant when you need it. Memories naturally fade over time unless they're important or frequently accessed, just like real memory.

Works with any LLM that supports MCP (Model Context Protocol)—Claude, GPT, Gemini, local models, and more.

An engram is a unit of cognitive information imprinted in a physical substance—the biological basis of memory.

What It Does

Tell your AI about your world:

"My colleague Sarah is allergic to shellfish and prefers window seats. She's leading the Q1 product launch."

Later, ask:

"I'm booking a team lunch and flights for the offsite—what should I know?"

Engram connects the dots—avoid seafood restaurants, book Sarah a window seat, and she's probably busy with the launch—and gives your AI the context to truly help.

It's not just search. It's understanding.

Quick Start

Install

npm install -g @199-bio/engram

Add to Your MCP Client

Claude Desktop — add to ~/Library/Application Support/Claude/claude_desktop_config.json:

{
  "mcpServers": {
    "engram": {
      "command": "npx",
      "args": ["-y", "@199-bio/engram"],
      "env": {
        "ANTHROPIC_API_KEY": "sk-ant-..."
      }
    }
  }
}

Claude Code:

claude mcp add engram -- npx -y @199-bio/engram

Other MCP clients — point to the engram command. It speaks standard MCP over stdio.

That's it. Your AI now has memory.

How Memory Works

Engram models memory like your brain does—important things stick, unimportant things fade, and everything connects.

flowchart LR
    subgraph Input
        A[Your Message] --> B[Remember]
    end

    subgraph Storage
        B --> C[Memory Store]
        B --> D[Knowledge Graph]
    end

    subgraph Retrieval
        E[Your Question] --> F[Recall]
        F --> G[Keyword Search]
        F --> H[Semantic Search]
        F --> I[Graph Traversal]
        G & H & I --> J[Combine Results]
    end

    C --> F
    D --> F
    J --> K[Relevant Memories]

Three Key Ideas

1. Memories Fade Over Time

Just like real memory, things you haven't thought about recently become harder to recall. But important or emotional memories resist fading.

Fresh memory (today)      → Easy to recall
Old but important memory  → Still accessible
Old, trivial memory       → Fades away

2. Accessing Memories Strengthens Them

Every time a memory is recalled, it becomes stronger and lasts longer. Frequently accessed memories become permanent.

3. Everything Connects

People, places, and things form a web of relationships. When you ask about Sarah, Engram also knows she works at Acme Corp and is leading the Q1 launch.

The Memory Lifecycle

flowchart TD
    subgraph "1. Capture"
        A[Conversation] --> B[Extract Key Info]
        B --> C[Store Memory]
        B --> D[Create Entities]
        B --> E[Link Relationships]
    end

    subgraph "2. Search"
        F[Query] --> G[Find by Keywords]
        F --> H[Find by Meaning]
        F --> I[Follow Connections]
        G & H & I --> J[Rank by Relevance + Recency]
    end

    subgraph "3. Consolidate"
        K[Raw Memories] --> L[Compress & Summarize]
        L --> M[Detect Contradictions]
        M --> N[Update Knowledge]
    end

    C --> F
    J --> K

Consolidation (Optional)

With an API key, Engram can periodically compress old memories into summaries—like how sleep consolidates your memories. This keeps important information while reducing storage.

# Run consolidation manually
consolidate  # Compresses old memories, finds contradictions

What Makes It Special

How to Use

Just talk naturally. Your AI will remember what matters.

Storing Memories

Say things like:

• "Remember that Sarah is allergic to shellfish"
• "My anniversary is March 15th"
• "I prefer morning meetings, never schedule anything before 9am"

The AI automatically extracts:

• Importance: Key facts get higher priority
• Entities: People, places, organizations mentioned
• Relationships: How entities connect to each other

Recalling

Just ask:

• "What do you know about Sarah?"
• "When is my anniversary?"
• "What are my meeting preferences?"

Results are ranked by:

• Relevance: How well it matches your question
• Recency: Recent memories surface first
• Importance: High-priority info stays accessible

The Knowledge Graph

Engram doesn't just store text—it understands relationships:

• People: Sarah, John, Dr. Martinez
• Places: Office, Conference Room A, Seattle HQ
• Organizations: Acme Corp, Stanford, FDA
• Connections: Sarah → works at → Acme Corp
• Observations: Sarah is allergic to shellfish

This means when you ask about Sarah's work, Engram knows to also surface relevant information about Acme Corp.

The Tools

Your AI gets these capabilities:

Examples

Personal Assistant

You: "Remember that my anniversary is March 15th and my partner loves Italian food"
AI: ✓ Stored. I've noted this as an important date with preferences.

You: "I need to plan something special next month"
AI: Your anniversary is coming up on March 15th! Given your partner's love for
    Italian food, perhaps a nice Italian restaurant reservation?

Relationship Context

You: "My friend David lives in Austin and always recommends great coffee shops."
AI: ✓ Stored. Created entity: David (person), with observation about Austin and coffee expertise.

You: "I'm visiting Austin next week"
AI: David is in Austin—he'd probably have great coffee shop recommendations!
    Want me to remind you to ask him?

Professional Memory

You: "John Chen is our VP of Engineering. He prefers async communication and hates meetings before 11am."
Claude: ✓ Stored. Created entity: John Chen (person, VP of Engineering).

You: "I need to schedule a sync with John"
Claude: Given John's preferences, I'd suggest a late morning slot, maybe 11:30am,
        or an async Loom video if it doesn't require real-time discussion.

Privacy

Your memories stay on your machine.

• All data stored locally in ~/.engram/
• No external APIs required for core functionality
• Consolidation uses Anthropic API (optional, requires key)
• You own your data completely

Technical Details

flowchart LR
    Q[Query] --> BM25
    Q --> Semantic
    Q --> Graph

    BM25 --> RRF[Rank Fusion]
    Semantic --> RRF
    Graph --> RRF

    RRF --> Decay[Apply Time Decay]
    Decay --> Salience[Weight by Importance]
    Salience --> Results[Final Ranking]

Retention = e^(-time / stability)

flowchart TD
    A[Old Memories] --> B{Important?}
    B -->|Yes| C[Keep as-is]
    B -->|No| D[Group Similar]
    D --> E[Summarize with AI]
    E --> F[Create Digest]
    F --> G[Archive Originals]

    H[All Memories] --> I[Find Contradictions]
    I --> J[Flag for Review]

engram/
├── src/
│   ├── index.ts              # MCP server entry
│   ├── storage/
│   │   └── database.ts       # SQLite + FTS5 + temporal fields
│   ├── graph/
│   │   └── knowledge-graph.ts
│   ├── retrieval/
│   │   ├── colbert.ts        # Semantic search
│   │   └── hybrid.ts         # RRF + decay + salience
│   ├── consolidation/
│   │   └── consolidator.ts   # Memory compression
│   └── web/
│       └── server.ts         # Visual browser

git clone https://github.com/199-biotechnologies/engram.git
cd engram
npm install
npm run build

# Install globally from local build
npm install -g .

pip install ragatouille torch

{
  "mcpServers": {
    "engram": {
      "command": "engram",
      "env": {
        "ENGRAM_DB_PATH": "/custom/path/",
        "ANTHROPIC_API_KEY": "sk-ant-..."
      }
    }
  }
}

Roadmap

• Core MCP server
• Hybrid search (BM25 + Semantic)
• Knowledge graph
• Entity extraction
• Temporal memory decay
• Memory consolidation
• Web dashboard
• Export/import
• Scheduled consolidation

Author

Boris Djordjevic Founder, 199 Biotechnologies

License

MIT — use it however you want.

Built with care by 199 Biotechnologies