Engram by EL AI IntelligenceEngram by EL AI Intelligence EL AI Intelligence EL AI Intelligence

Your AI deserves a memory.

One vault, every tool. Claude, Cursor, Windsurf, any MCP client, your browser, your CLI — your memory follows you across platforms, not tools. And privacy you can verify, not just trust: encryption runs on your machine, sync servers hold only ciphertext, and the storage format is open (Apache-2.0) — your memories stay yours, and stay readable.

⬇ Install in one command Open Vault →
$ curl -fsSL https://engram.ellmstack.dev/install.sh | bash
$ engram onboarding  # vault + first memory in ~5 minutes
Uninstall: curl -fsSL https://engram.ellmstack.dev/uninstall.sh | bash
$ curl -fsSL https://engram.ellmstack.dev/install.sh | bash
$ engram onboarding  # vault + first memory in ~5 minutes
Uninstall: curl -fsSL https://engram.ellmstack.dev/uninstall.sh | bash
> irm https://engram.ellmstack.dev/install.ps1 | iex
> engram onboarding  # vault + first memory in ~5 minutes
Windows binaries are unsigned — the installer fetches them directly, so SmartScreen stays quiet. Direct browser downloads need More info → Run anyway. Uninstall: irm https://engram.ellmstack.dev/uninstall.ps1 | iex
Package manager: choco install engramd (Chocolatey)
The wider world
Built for AI. Not only for AI.
Engram's vault is aimed at AI agents today — but the problem it solves is far older than AI. Humans, teams, and software have been deciding what to keep, what to strengthen, and what to forget since long before the first prompt. The model of memory underneath doesn't care who's doing the remembering.
Also from EL AI Intelligence: Guardrail — the firewall for AI agents that act · Amparo — an open agent that acts under policy
How it works
Three steps to never forgetting again.
Engram by EL AI Intelligence runs as a memory daemon on your machine. Everything talks to it through one JSON API — MCP clients, Python/JS SDKs, chat bots, the browser extension — and so can anything else that needs a memory. If your stack can POST JSON, it has memory now.
1

Install

One command — Linux/macOS: curl -fsSL https://engram.ellmstack.dev/install.sh | bash — then engram onboarding: encrypted vault, first memory, and a running daemon in about five minutes. (Windows uses the same one-liner: irm https://engram.ellmstack.dev/install.ps1 | iex — unsigned, so it installs via the script, where SmartScreen doesn't fire.)

2

Connect

engram mcp install wires Claude Desktop, Cursor, Windsurf, and Claude Code to your vault in one command. Capture, search, and context become tools your AI just has.

3

Recall

Engram by EL AI Intelligence assembles relevant context for every query, so your LLM's token cost stops growing as your memory grows. Forgetting is efficient — and green.

MCP everywhere
One memory. Every AI you use.
One command — engram mcp install — writes an engram entry into every supported editor's MCP config, merging with whatever servers you already have. And the same vault serves any MCP client, the browser vault (read-only unlock from anywhere), a browser extension, Slack, Discord & Telegram bots, and the CLI. Capture once in Claude Desktop, recall it in Cursor, browse it in your browser.

Claude Desktop

Config merged into claude_desktop_config.json — no editing by hand, nothing overwritten.

Claude Code

One command: claude mcp add engram -- engramd-mcp

Cursor & Windsurf

Wired via ~/.cursor/mcp.json and ~/.codeium/windsurf/mcp_config.json — same one command.

Slack, Discord & Telegram

Capture team decisions right from chat — mention the bot, /remember, react with 📌 (Slack & Discord), or watch a channel. Self-hosted bot first — keys never leave your machine.

Browser extension

Capture the page you're on — your selection, the title, or the link — and search your vault from the popup. Chrome, Edge, Brave & Firefox (121+).

Personal

$10 /month, flat

End-to-end encrypted sync on 5 devices · 10 GiB storage.

Get Personal →

Team

$29 /month, flat

Shared team memory: 25 devices · 100 GiB · teams up to 10 members. No per-seat charges.

Get Team →

Organization

$99 /month, flat

Company-wide memory: 50 devices · 500 GiB · unlimited members. No per-seat charges.

Get Organization →

End-to-end encrypted sync included — the relay stores only ciphertext. Free tier: 1 device, 1 GiB. Self-hosting stays free.

The science
Memories that learn to forget.
Engram models human memory: Ebbinghaus decay curves, holographic XOR codes for O(1) associative recall, and nightly consolidation that merges episodes into semantic knowledge. And deletion is honest: delete a memory and it is gone — from search, from sync, from the relay.

Ebbinghaus Forgetting Curve

100% 0% 30d Time Strength
Without retrieval With spaced retrieval

Memories you revisit get stronger. Memories you ignore fade away. Engram's decay model mirrors how your own brain decides what to keep.

Deletes that actually delete.

  • ✓Atomic. Row, search index, and tombstone are written in one transaction — no window where a deleted memory lingers.
  • ✓Sync-aware. The tombstone rides the sync relay, so every device forgets together — and the relay physically erases the ciphertext. Gone means gone, everywhere.
  • ✓Auditable. A local access ledger records every read and delete — you can see who touched what, even after deletion.
Environmental impact
Forgetting is green.
Every memory retrieved is context the model doesn't have to re-create. We measured the trade-off against the alternative: an A/B benchmark pitting Engram's retrieval against wholesale context injection on the same corpus.

+4.5% vs +11%

When the memory corpus grew 4× (25→100 memories), Engram's input tokens per question grew +4.5% — bounded by top-k retrieval. Wholesale injection grew +11% and rising: linear by construction, and it fills the entire context window around ~4,000 memories.

Embedded once, not rebroadcast

Injection sends the whole knowledge base into every interaction, in every session, forever. A 20-person team asking 10 questions a day that touch memory, at 1,000 memories, burns ≈14M vs ≈10M input tokens/day — and every 1,000 memories added costs injection another +8.4M/day while Engram barely moves.

Retrieval runs locally

Embeddings and search run on your machine — inference stays local. The only datacenter load per query is the tokens your AI actually reads. Hosted RAG pays for retrieval compute too; Engram doesn't.

4h
20,000
50%
Estimated annual CO₂ savings
—
kg CO₂e per year
Based on 0.4 g CO₂e per 1K tokens (public estimate). Your vault dashboard shows the same estimate for your own activity.

Measured, not vibes — and honest about where it doesn't hold: below ~100–150 memories, retrieval costs a few percent more tokens.

Compare
Why Engram by EL AI Intelligence?
Most memory solutions are vector databases with a search bar. Engram by EL AI Intelligence is a memory system — with forgetting, consolidation, E2E sync, explicit MCP capture, and deletes that actually delete.
Feature Engram by EL AI Intelligence Mem0 ChromaDB Letta
Local-first ✓ SQLCipher ✓ Self-host ✓ Embedded ✗ Server req.
Ebbinghaus decay ✓ Built-in ✗ ✗ ✗
Holographic cache ✓ QEM L1 ✗ ✗ ✗
Nightly consolidation ✓ Episodic→Semantic ✗ ✗ ✓ Agent memory
MCP capture (explicit) ✓ On call — never automatic ✗ ✗ ✗
E2E encrypted sync ✓ Shipped ✗ ✗ ✗
Deletion audit trail ✓ Access ledger ✗ ✗ ✗
CO₂ tracking ✓ Instrumented ✗ ✗ ✗
SDKs Python, JS, MCP Python Python, JS Python, REST
License Open-core · format Apache-2.0 Apache 2.0 Apache 2.0 Apache 2.0

Built with Rust. Runs on your machine. Your memories, your key.

Zero telemetry. No cloud required. With sync, encryption happens on your machine — servers hold only ciphertext. And the format that reads your memories is open source, so retrieval never depends on this product.

Open format · GitHub Format spec · FORMAT.md Open Vault →