Memory AI

Engram's /memory page is a live demo of an AI that never forgets. Every message you send is permanently stored as a vector embedding on the Engram decentralized network, recalled by semantic similarity across all your future sessions.

How it works

Every conversation turn goes through four steps before the AI replies:

Your identity — no login required

Engram uses a stable anonymous UUID stored in your browser's localStorage as your identity. No account, no email, no wallet needed.

// Your ID lives here — never leaves your browser
localStorage.getItem("engram_session")
// → "s_m3x7k2_abc9f1..."

This ID is used to:

• Tag every message stored on Engram with your session
• Filter recalled memories so you never see another user's data
• Load your full chat history from the server on any new visit
• Generate shareable read-only links to your conversations

Chat history persistence

Messages are saved in two places simultaneously so your history is never lost:

When you load the page, the server copy wins if it has more messages than your local cache — for example after visiting from a different device.

Cross-device sync

To continue your conversation on a different device or browser, copy your session ID from the original browser and paste it into the new one:

// On your original browser — open DevTools console:
localStorage.getItem("engram_session")
// Copy the result, e.g. "s_m3x7k2_abc9f1..."
// On your new browser — paste it:
localStorage.setItem("engram_session", "s_m3x7k2_abc9f1...")
// Reload the page — your history loads from the server.

Sharing a conversation

Click the Share button in the top-right of the chat to copy a read-only link. The full conversation is base64-encoded into the URL — no server needed, works anywhere.

# Share URL format
https://theengram.space/memory?view=eyJyIjoie...
# Anyone who opens this sees a read-only replay of the chat
# with a "Start your own" button to begin their own session

Chat history folders

The chat interface groups your messages into date sections — Today, Yesterday, and then by calendar date for older conversations. Each section is collapsible. This makes it easy to find a specific conversation without scrolling through everything.

Build your own memory-powered AI

The full stack is open source. The key pieces are:

1. Store a message on Engram

from engram.sdk import EngramClient
client = EngramClient("http://your-miner:8091")
# Store a user message with session metadata
cid = client.ingest(
    "I'm 22 years old and I built Engram",
    metadata={
        "role": "user",
        "session": "s_m3x7k2_abc9f1",
        "text": "I'm 22 years old and I built Engram",
    }
)
print(cid)  # v1::a3f9...

2. Recall memories before generating a reply

# Query for semantically related past messages
results = client.query("how old are you", top_k=30)
# Filter to this user's session only
session_memories = [
    r for r in results
    if r["metadata"].get("session") == "s_m3x7k2_abc9f1"
    and r["metadata"].get("text")
][:12]
# Build context
memory_lines = [
    f"[{r['metadata']['role'].title()}]: {r['metadata']['text']}"
    for r in session_memories
]
context = "\n".join(memory_lines)

3. Inject into your LLM call

import anthropic  # or openai, xai, etc.
client_ai = anthropic.Anthropic()
system = f"""You are a memory-powered AI assistant.
You have access to the user's past conversation history:
{context}
Treat these as real facts. Never deny knowing something that appears above."""
response = client_ai.messages.create(
    model="claude-haiku-4-5-20251001",
    max_tokens=1024,
    system=system,
    messages=[{"role": "user", "content": user_message}]
)

4. Store the AI response too

reply_text = response.content[0].text
# Store so it can be recalled in future turns
cid = client.ingest(
    reply_text,
    metadata={
        "role": "assistant",
        "session": "s_m3x7k2_abc9f1",
        "text": reply_text[:500],
    }
)

Privacy & data ownership

• Session isolation: Engram returns up to 40 candidate memories per query but only those matching your session ID reach the AI. Other users' memories are discarded server-side before context injection.
• Decentralized storage: Messages are stored as vector embeddings across Bittensor miners — not in a central database controlled by Engram or any single entity.
• No PII required: Your identity is an anonymous UUID. No email, name, or wallet address is ever requested or stored.
• Sensitive data: For production use cases involving sensitive information, use Private Namespaces which encrypt your data with a PBKDF2-derived key before storing on miners.