aikito

What Aikito governs

Aikito is a Git-managed workspace for governing AI agent context and durable memory. The README states the formula plainly: Aikito equals governing agent resources multiplied by curating durable memory. Plain files define the source of truth, explicit scopes define who sees what, and Git keeps the history. The problem it addresses is that agent resources fragment across tools, across projects, and across time. Each agent tends to need a different configuration format. Reusable knowledge, skills, and instructions get copied or maintained across multiple repositories. Decisions and lessons disappear into old sessions. Aikito keeps all of it in one personal Git workspace and exposes selected resources to each agent and project. The workspace layout has top level directories for skills, memory, global instructions, shared MCP definitions, and shared subagents, plus a projects directory where each project gets its own agent.toml, AGENTS.md, and memory directory. The README emphasizes that no database, daemon, vector store, or hosted service is required. The resource table maps each kind, such as memory or skills, from its canonical source to its synchronized destination, which includes both global access and per project runtime locations. The default registry of supported agents includes Codex, Claude Code, Antigravity CLI, OpenCode, GitHub Copilot CLI, and the DeepSeek Harness.

How memory and sync work

The durable memory half is a bundled skill that guides coding agents to retrieve relevant notes before acting, distill durable conclusions from what they learn, update notes that went stale, and choose the right global or project scope. New workspaces enable this workflow by default: initialization creates the memory structure, installs and selects the bundled aikito and durable-memory skills, and adds the minimal global instruction that requires agents to evaluate memory relevance. Nothing is connected to an agent until the user explicitly runs aikito sync global. The notes are ordinary Markdown, so Git gives history, review, rollback, and portability, and the memory an agent wrote in one client is the same memory another client reads later. A user can share a resource with link mode, which keeps it shared and immediately up to date, or copy mode, which gives a project an isolated snapshot it can evolve independently. Project memory always uses link mode with its canonical scope, preserving one history. The command aikito show memory lists what has accumulated, grouped by scope, and aikito maintain memory runs confirmation gated, full scope maintenance. A local read-only web console bound to 127.0.0.1 lets a user browse the workspace, resources, scopes, and governance state without changing anything. The project deliberately does not capture every agent action, run a vector store, inject context through a background daemon, orchestrate supervisor and worker agents, or replace a coding agent's native runtime.

Requirements and setup

Aikito requires macOS or Linux, with Windows users pointed to WSL2, because it relies on symbolic links and POSIX file permissions for synchronization and credential safety. The supported Python versions are 3.12, 3.13, and 3.14, and Git is required. The recommended install is through Homebrew with brew install lsaint/tap/aikito, which also sets up shell completion. The manual flow is aikito init workspace to create the Git repo, aikito sync global to synchronize global resources, and aikito status to report state across supported agents. For each code project, aikito init project registers it and connects its resources to the project's agent configuration. The status command prints a table showing which agents have instructions, skills, MCP config, and subagents synced. An adopt command imports an existing setup and previews changes read only before applying, then creates timestamped backups and imports detected configurations without overwriting originals. A companion project, Chat Distiller, turns browser AI conversations into reviewable Markdown notes and saves them to the Aikito inbox. The safety section warns that initializing a workspace creates a local Git repository that is not private by default, and that deleting a later commit does not remove a secret from history, so a user should review memory and configuration before adding a remote. Contributors run the unit tests with python3 -m unittest discover before submitting code, and vulnerabilities are reported privately under the project security policy.

Editorial conclusion

The repository is published under the MIT license and its most recent commit was recorded on 2026-08-26. The source lives at https://github.com/lsaint/aikito.