Ambient Memory: the local-first context layer Jarvis Osorio built for AI engineers
By Jarvis Nelson Osorio · November 16, 2026
Every AI coding session starts with the same tax: re-explaining what you were doing. Which file you were editing, which doc you had open, which page you read twenty minutes ago. Ambient Memory, an open-source project by Jarvis Osorio (Jarvis Nelson Osorio), is his answer to that tax. It is a local-first context layer that passively remembers what you worked on — files, screen sessions and browser history — and exposes that memory to Cursor and Claude Desktop over the Model Context Protocol (MCP).
What it is, in plain terms Ambient Memory is written in TypeScript, requires Node.js 20 or later (22+ preferred) and is released under the MIT License. The current version in the repository is 0.3.0, and there is no telemetry or network sync in v0.3. Everything stays on your machine.
A background daemon watches Documents, Desktop, Downloads and the current working directory, and indexes screen and browser activity with noise filters. It ignores apps such as Slack and Spotify and dedupes focus churn, so the memory reflects work rather than every window switch.

An architecture built around one owner of the data The design choice Jarvis Osorio highlights most is that only the daemon opens the SQLite vault. The CLI and the MCP server are thin clients over a local Unix socket (or a Windows named pipe); they never open the database themselves. That single-owner pattern keeps locking simple and keeps the attack surface small.
Storage uses Node's built-in node:sqlite on Node 22+, with a sql.js WASM fallback on Node 20. There is no better-sqlite3 native addon, so there is no node-gyp or MSVC toolchain to fight during install. Search uses FTS5 with BM25 ranking.
Privacy is treated as an engineering problem, not a slogan. Entity fields are encrypted with AES-256-GCM. When the daemon stops, the database is sealed to memory.db.enc and the plaintext plus WAL and SHM files are removed. The project is also honest about its tradeoff: while the vault is unlocked, the FTS index is redacted plaintext so that search can work. That is documented in the repository rather than hidden.
The MCP tools The MCP server auto-starts the daemon if needed and exposes six tools:
- memory_search — full-text search with recency and path boosts, plus type, path and time filters
- memory_recent — recently seen entities and ambient events
- memory_get — fetch one entity by id, including links
- memory_context — a narrative "what was I doing" summary for the last N minutes
- memory_relate — create or update a typed relation between two entity ids
- memory_status — a health and privacy snapshot: store counts, integrity, key backend and daemon state
memory_context is the one that changes the feel of a session: instead of pasting context into a chat, the agent can ask what you were doing in the last half hour.

Getting started The package is not on the npm registry yet, so install from GitHub:
git clone https://github.com/JarvisOsorio/ambient-memory.git
cd ambient-memory
npm install
npm test
node ./bin/ambient.js init
node ./bin/ambient.js start --screen auto
node ./bin/ambient.js mcpThen point Cursor or Claude Desktop at the repo binary in your MCP config, using "command": "node" and the absolute path to bin/ambient.js with the "mcp" argument. The CLI also offers status, doctor, search, recent, context and forget — including forget --all --yes when you want a clean slate.
Why it matters Most "memory" features for AI assistants live in someone else's cloud. Jarvis Osorio took the opposite route: a memory that is local, encrypted at rest, inspectable, and deletable on command. For engineers who want agents with context but not surveillance, that is the point.

Read the full project page on jarvisosorio.com and browse the source on GitHub. Jarvis Osorio also wrote about the idea in his LinkedIn article "Ambient Memory: Why I Built a Persistent State Layer for Autonomous Agents" (November 12, 2026).