One memory layer for every MCP-compatible agent. Runs on your machine. Yours forever — no re-explaining your codebase, no re-uploading your notes.
Capture in Claude Code on your laptop. Recall in Cursor on your desktop. Same brain, no copy-paste.
Six primitives, one local SQLite file. Plug it into any MCP-compatible runtime and let your agent remember — across sessions, sources, and machines.
21 tools exposed over the Model Context Protocol. Drop-in for Claude Code, Cursor, OpenAI agents, custom runtimes — no SDK, no wrapper.
// claude_desktop_config.json
{
"mcpServers": {
"engram": {
"command": "engram-mcp",
"args": ["--port", 8765]
}
}
}SQLite + LanceDB. Embeddings via Ollama. Nothing leaves your machine unless you say so.
Drive · Notion · YouTube · Obsidian · voice · agent transcripts.
Your agent supplies titles and tags. We add vectors and views.
Voice, text, link share — capture from anywhere, sync back home.
Bidirectional ops-log over XChaCha20-Poly1305. Your laptop and desktop share one brain. We see ciphertext, never content.
Browse your engram visually. Nodes pulse, synapses connect — click to read, hover to recall the trace.
Every other memory layer is passive — it stores and recalls. Engram is the only one whose tools let the agent flag what's missing from the graph and surface patterns across captures.
find_gaps— your agent asks the question you forgot to answer.analyze_patterns— clusters and trends across thousands of memories, on demand.summarize_recent— a daily digest your agent can pull on its own.brew install to recall in 3 steps.No cloud account required to get started. The local server runs as a daemon and exposes itself to any MCP client.
The Engram daemon installs locally, sets up SQLite + LanceDB, and downloads the Ollama embedding model in the background. Works on macOS, Linux, Windows (WSL).
Add one block to your MCP config. Claude Code, Cursor, Cline, Continue — they all discover Engram's 21 tools automatically on next launch.
Every meaningful exchange gets a memory. Every recall pulls semantic context from the local vector store. The graph grows over time — yours, queryable, exportable.
Memory layers exist. Most live in someone else's cloud, speak only one agent, and ship as a closed SDK. Here's how Engram differs.
✓ supported · ~ partial / via plugin · ✗ not supported · — not applicable
Each tool is a single verb your agent can call. Self-documenting via MCP — your agent introspects them automatically.
Local-first is the default. The Pro tier adds exactly two cloud-side primitives — sync transit and hosted compute. Everything else stays on your disk. Here's the full data-handling ledger.
The local server is MIT-licensed. Pay only when you want mobile capture, multi-PC sync, and an always-on dashboard.
Billing via Polar (Merchant of Record — handles EU VAT). Cancel anytime from your dashboard. No refunds for partial months.
In neuroscience, an engram is the physical trace a memory leaves in cortical tissue — a small population of neurons that fires together when a memory is recalled. We borrowed the metaphor: every capture you make is a node, every connection a synapse, and recall is a pattern firing across the graph.
No. The local server runs entirely offline. Embeddings via Ollama, transcription via Whisper, vector search via LanceDB — all local. You only need to be online if you want to pull from cloud sources (Drive, Notion) or use the Pro sync.
Yes. The memory store is a regular SQLite file at ~/.engram/memory.db. You can copy, backup, or open it with any SQLite tool. We also expose an export CLI command that writes Markdown + JSON.
The MCP server (the local daemon, all 21 tools, ingestion pipelines) is MIT-licensed and available on GitHub. The cloud workers (sync, hosted AI) that power the Pro tier are proprietary — they’re what we charge for.
Your master key is derived from a passphrase via Argon2id. It lives only on your devices. Anything that transits our cloud is encrypted with XChaCha20-Poly1305 client-side, before upload. Server-side, we see ciphertext — never content. The dashboard decrypts in-browser, using a key sent over your local pairing channel.
Anything that speaks MCP: Claude Code, Cursor, Cline, Continue, Zed (via plugin), custom runtimes using the official MCP SDKs. The 21 tools are exposed identically across all of them.
The $9 includes 10M embedding tokens and 5h of Whisper transcription per month. Overages are billed at $0.05 per million embedding tokens and $0.20 per hour of Whisper. You can also bring your own keys (Voyage, OpenAI) and skip the hosted compute entirely.
One command and your Claude Code, Cursor, and custom agents start sharing what they learn — for free, local-first, on the box you already own.