Engram

Give your AI a perfect memory.

Every conversation you have with your AI disappears the moment it ends. Names you've mentioned, preferences you've shared, the context of your life—all gone. You find yourself repeating the same information, re-explaining who people are, reminding it of things you've already said.

Engram changes that.

It gives your AI the ability to remember. Not just store text, but truly remember—the way you do. Important things stick. Trivial things fade. And everything connects to everything else.

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

How It Works

Tell your AI something once:

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

Weeks later, ask:

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

Engram connects the dots. It remembers Sarah, her allergy, her seating preference, and her current workload. Your AI can now actually help—suggesting restaurants without shellfish, booking her a window seat, and noting she might be busy with the launch.

This isn't keyword matching. It's understanding.

Memory That Feels Real

Engram models memory the way your brain does.

Things fade. A memory from six months ago that you've never thought about becomes harder to find. But something important—a name, a birthday, a preference—stays accessible even as time passes.

Recall strengthens. Every time a memory surfaces, it becomes more permanent. The things you think about often become the things you'll never forget.

Everything connects. People link to places. Places link to events. When you ask about one thing, related things come along for the ride. Ask about Sarah, and her company, her projects, and her preferences all surface together.

Quick Start

Install globally:

npm install -g @199-bio/engram

Add to Claude Desktop (~/Library/Application Support/Claude/claude_desktop_config.json):

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

Or with Claude Code:

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

That's it. Your AI now remembers.

What You Can Do

Just talk naturally. Your AI handles the rest.

Store memories by mentioning things:

• "Remember that my anniversary is March 15th"
• "Sarah prefers async communication"
• "I'm allergic to penicillin"

Recall memories by asking:

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

Build a knowledge graph of your world:

• People, places, organizations, and how they connect
• Observations about each entity
• Relationships that span your entire life

Privacy

Your memories stay on your machine. Everything is stored locally in ~/.engram/. The only external call is optional—if you provide an API key, Engram can periodically compress old memories into summaries. But the core functionality works entirely offline.

The Details

Your AI gets these capabilities:

| Tool | Purpose | |------|---------| | remember | Store new information with importance, emotions, and timing | | recall | Find relevant memories ranked by relevance and recency | | forget | Remove a specific memory | | create_entity | Add a person, place, or concept to the knowledge graph | | observe | Record a fact about an entity | | relate | Connect two entities (e.g., "works at", "married to") | | query_entity | Get everything known about someone or something | | list_entities | See all tracked people, places, and things | | stats | View memory statistics | | consolidate | Compress old memories and detect contradictions | | engram_web | Launch a visual memory browser |

Engram uses three search methods simultaneously:

1. Keywords — SQLite FTS5 finds exact matches for names, dates, and phrases
2. Meaning — Neural embeddings find conceptually related content
3. Connections — The knowledge graph expands to related entities

Results are fused together, then adjusted for how recent and important each memory is. Fresh memories surface first. Important memories resist fading.

Memories follow an exponential decay curve:

Retention = e^(-time / stability)

• Time is days since the memory was last accessed
• Stability is memory strength, which increases each time you recall something

High-importance and emotionally significant memories decay slower. Frequently accessed memories become essentially permanent.

With an API key, Engram can compress old memories—like how sleep consolidates your experiences into long-term storage.

1. Groups related low-importance memories together
2. Creates AI-generated summaries (digests)
3. Detects contradictory information
4. Archives the originals

This keeps storage efficient while preserving everything important.

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

Environment variables:

| Variable | Purpose | Default | |----------|---------|---------| | ENGRAM_DB_PATH | Where to store data | ~/.engram/ | | ANTHROPIC_API_KEY | Enable consolidation | None (optional) |

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

For semantic search, install Python dependencies:

pip install ragatouille torch

If unavailable, Engram falls back to keyword-only search automatically.

On M1 MacBook Air:

| Operation | Time | |-----------|------| | Remember | ~100ms | | Recall | ~50ms | | Graph queries | ~5ms | | Consolidate | ~2-5s per batch |

Storage: ~1KB per memory.

Roadmap

• [x] Hybrid search (keywords + semantics)
• [x] Knowledge graph with relationships
• [x] Memory decay and strengthening
• [x] Consolidation with contradiction detection
• [x] Web interface
• [ ] Export and import
• [ ] Scheduled consolidation

Author

Boris Djordjevic Founder, 199 Biotechnologies

License

MIT