What it does

piia-engram is an open-source, local-first layer that holds an AI work identity for you: your preferences, quality standards, lessons learned, key decisions and project context. MCP-compatible coding tools such as Claude Code, Codex, Cursor and Windsurf can read the same approved context, so a new chat or a different tool starts from what you have already told it. The README says it is not an agent memory database but a user-owned layer above your tools. Data lives under ~/.engram as JSON and Markdown files you can open and edit, with no cloud account. The AI proposes knowledge; higher-risk items such as credentials, shell commands, MCP config and permission rules wait for your review, while other writes are absorbed but stay auditable and reversible. A preview command called Memory Lens shows what an AI caller would receive and what was withheld. You install it with pip and run engram setup, whose wizard detects supported tools, lists the config files it will touch and writes the MCP connection after you confirm. It exposes 19 core MCP tools by default and 40 more as an opt-in, and it can export a Markdown identity card for tools without MCP, such as ChatGPT and Gemini. By default it sends one anonymous usage ping a day, which can be turned off. The code is under the GNU AGPL v3.

Capabilities

Stores identity, preferences, lessons, decisions and playbooks as local JSON and Markdown files
MCP server for Claude Code, Codex, Cursor, Claude Desktop and other MCP clients
Review gate for high-risk writes, with auditable and reversible changes
Memory Lens preview of what an AI caller would receive
Markdown identity card for tools without MCP support
Optional hybrid search combining full-text and semantic retrieval

When to use it

You switch between several AI coding tools and keep re-explaining the same preferences
You want AI memory kept in files you own rather than inside one vendor's product
You want lessons and decisions from past projects available when you start a new one