Engram

One shared memory for all your AIs.
Your assistants forget you the moment a chat ends — and each one lives in its own silo. Engram is the memory they share.

What it is

Pull your context in from any AI — a ChatGPT or Claude export, any Markdown / text / JSON, or a quick paste — into one private, local memory. Then any assistant that speaks MCP (Claude Desktop, Claude Code, your own apps) can recall it, across every conversation you've fed in. Remember once; every connected AI knows it.

You own all of it: memories are plain Markdown files on your Mac (memories/<conversation>.md) — readable, greppable, editable, deletable. Nothing is ever uploaded.

How it works

• One memory, many minds. A single store every MCP client reads from and writes to. "Collective mind" recall (.all) searches across all your conversations at once; or scope recall to a single conversation, or to one named file you've grouped.
• Bring your history in. Import a ChatGPT / Claude export, Markdown/text/JSON, or paste — and choose Key facts (keep only the durable things you said) or Everything. Auto-capture from Claude Code transcripts is built in (opt-in).
• Semantic recall, on device. Embeddings run locally via Apple's Natural Language framework (an optional one-time download upgrades to the contextual transformer). No model server, no network for recall.
• Native, and yours. A real native macOS app (SwiftUI + AppKit — no Electron, no web wrapper). Local-first by design: no account, no telemetry, no analytics. The only network is opt-in and never sends your data — a manual "Check for Updates", an optional local-Ollama distiller, and the one-time on-device model download. See PRIVACY.md.

Install

Homebrew — the MCP server + capture CLI, the fastest path:

brew install albertofettucini/engram/engram
engram-mcp --connect                 # register Engram with Claude Desktop, then restart it
engram-mcp --prepare-embeddings      # optional: one-time on-device model for better recall
engram-capture --watch               # optional: auto-capture durable facts from Claude Code

The desktop app — to browse, edit, and import your memories: grab Engram-<version>.zip from the latest release, unzip, drag it to Applications, then right-click → Open the first time. The app is unsigned (no paid Apple cert), so Gatekeeper asks once; after that it opens normally.

Connect any MCP client. Anything that speaks MCP reads and writes the same store. --connect writes the Claude Desktop config for you; to wire another client by hand, point it at the engram-mcp binary (which engram-mcp — usually /opt/homebrew/bin on Apple Silicon, /usr/local/bin on Intel):

{
  "mcpServers": {
    "engram": { "command": "/opt/homebrew/bin/engram-mcp" }
  }
}

Build from source — no Homebrew needed:

git clone https://github.com/albertofettucini/Engram && cd Engram
swift build -c release
.build/release/engram-mcp --connect
bash packaging/make-app.sh           # → Engram.app on your Desktop

Requirements: the app needs macOS 14+; the CLI tools (MCP server, capture) run on macOS 13+. No backend, no third-party services — the engine, MCP server, and capture tool are pure Swift + Foundation.

Engine API

let engine = try MemoryEngine(root: memoriesFolderURL)
try engine.remember("I prefer dark mode.", source: "manual", conversation: "chat-1")
let hits = engine.recall("what theme do I like", scope: .all)   // collective mind
let where_ = engine.search("dark mode")                         // keyword → which conversation
try engine.forget(hits[0].memory.id)                            // soft delete (stays on disk)

Build & test

swift build
swift test

More from the same maker

Council — ask several LLMs one question; they answer in parallel, peer-review each other blind, and you see exactly where they disagree.

License

MIT © 2026 Joseph.

_{Your AIs forget. Engram remembers — and shares it.}