pacifio/atlas

What it solves

Atlas provides source control for coding agents, solving the problem of "lost reasoning" where agents write code but the context, prompts, and tool calls that led to that change are not preserved. It also addresses the fragmentation of agent memory, allowing different agents (like Claude Code and Codex) to share a persistent, on-device memory of decisions and plans so that switching agents mid-task doesn't require starting over.

How it works

Atlas acts as an enrichment layer that runs agents as subprocesses (via the Agent Client Protocol - ACP) or in-process. Before a prompt is sent, Atlas assembles a context pack consisting of locally resolved @ mentions, shared agent memory, semantic matches from a local embedding index, and session handoffs.

It records every session to a local SQLite database. When a commit is made, Atlas links that commit back to the specific agent session that produced it, creating a "checkpoint" that preserves the reasoning behind the change.

Who it’s for

Developers who use AI coding agents and want a permanent, searchable record of why changes were made, as well as a unified workspace where multiple different agents can collaborate on the same codebase with shared memory.

Highlights

• Checkpoints: Links Git commits to the agent sessions, prompts, and reasoning that produced them.
• Shared Memory: An on-device semantic index that allows different agents to share plans, architecture notes, and decisions.
• Multi-Agent Support: Run Claude Code, Codex, and other ACP-compatible agents side-by-side in one window.
• Local-First: Code, notes, and sessions are stored locally by default, with secrets scrubbed before persistence.
• Integrated Workspace: Includes a built-in editor, terminal, Git graph, and a markdown-based knowledge base.