personal-model Builds Your AI a HUMAN.md So It Stops Forgetting Who You Are

Switch between Claude Code, Codex, and Cursor in the same afternoon and every one of them starts cold: no idea what you're working on, what you already decided, or how you like things done. personal-model is a local-first tool built to fix exactly that, by turning your own Mac activity into a single portable file, HUMAN.md, that any MCP-compatible agent can read.

What it actually builds

After you grant macOS permission, personal-model learns from focused activity captured on your machine and structures it into five progressively deeper layers, described in the project's own docs: a Point is a single sourced observation, a Line is a relationship or change over time, a Face is a pattern backed by related evidence, a Volume is a higher-order structure spanning projects or areas of your life, and the Root is the current integrated model of you. The project's own numbers for a mature model: over 1,000 Points, 300-plus Lines, 80-plus Faces, and 20-plus Volumes, distilled into one Root. Every claim in that structure keeps its source receipt; new evidence can strengthen, revise, or overturn an earlier inference rather than silently overwriting it.

The pitch that separates this from a generic chat-history feature: the same prompt to the same AI should not produce the same answer for two different people, because two people have different goals, different working patterns, and different standards for what "done" looks like. Connecting a Personal Model changes how a connected agent prioritizes, decides, and writes, because it understands who it's actually working for.

How you connect it

Installation is a package-managed CLI: uv tool install personal-model, then persome onboard and persome model open --after 30 to review the first pass at your model. The distribution is named personal-model, the CLI itself is persome. From there, it connects to whichever agents you actually use: persome install claude-code for Claude Code, persome install codex for Codex CLI and its IDE extension, persome install cursor-agent for Cursor Agent CLI, plus managed configs for Claude Desktop, opencode, and any other MCP-compatible client. One Personal Model, multiple agents reading the same evidence-linked context instead of each one starting over.

The project draws a specific line worth understanding before you connect anything: personal-model is an MCP server used by trusted MCP clients, not a general integration hub. Other MCP servers, Filesystem, GitHub, Slack, Google Drive, are separate tools, not personal-model integrations, unless that specific path has been explicitly built and tested.

What it's actually for, day to day

The concrete use case the project leads with is a connected coding agent that finds its own work: searching the Personal Model for unfinished tasks, ranking proposed next steps against your actual priorities, and separating what it can just do locally from external actions that still need your explicit approval. Two workflows the docs describe directly: picking up a project exactly where you left off, restoring the current goal, the next step, and the git status without you re-explaining any of it, and a version that runs while you're away, reviewing open loops and preparing a report by morning, again with anything external gated behind your approval rather than acted on autonomously.

Where the honesty lives

Two things are worth knowing before you install this rather than after. First, it requires macOS 13 or newer plus Xcode Command Line Tools, this is a Mac-only tool right now, not cross-platform. Second, and more important: it explicitly requires you to grant macOS permission for it to observe your activity in the first place, and everything downstream depends on trusting that boundary. The project's counter to that is that the model is private by default, runs entirely locally, and is "yours to inspect, correct, export, and delete," not dependent on a hosted memory service, but that's a design commitment to evaluate on its own terms, not a reason to skip reading what permission you're actually granting.

Free and open source under the Apache 2.0 license, built by Intuition-Lab, a newer entrant (created July 2026) already past 1,200 GitHub stars as of this writing.