diff --git a/ARCHITECTURE.md b/ARCHITECTURE.md
new file mode 100644
index 000000000..f93603137
--- /dev/null
+++ b/ARCHITECTURE.md
@@ -0,0 +1,88 @@
+# Cargo Architecture
+
+This document gives a high level overview of Cargo internals. You may
+find it useful if you want to contribute to Cargo or if you are
+interested in the inner workings of Cargo.
+
+
+## Subcommands
+
+Cargo is organized as a set of subcommands. All subcommands live in
+`src/bin` directory. However, only `src/bin/cargo.rs` file produces an
+executable, other files inside the `bin` directory are submodules. See
+`src/bin/cargo.rs` for how these subcommands get wired up with the
+main executable.
+
+A typical subcommand, such as `src/bin/build.rs`, parses command line
+options, reads the configuration files, discovers the Cargo project in
+the current directory and delegates the actual implementation to one
+of the functions in `src/cargo/ops/mod.rs`. This short file is a good
+place to find out about most of the things that Cargo can do.
+
+
+## Important Data Structures
+
+There are some important data structures which are used throughout
+Cargo.
+
+`Config` is available almost everywhere and holds "global"
+information, such as `CARGO_HOME` or configuration from
+`.cargo/config` files. The `shell` method of `Config` is the entry
+point for printing status messages and other info to the console.
+
+`Workspace` is the description of the workspace for the current
+working directory. Each workspace contains at least one
+`Package`. Each package corresponds to a single `Cargo.toml`, and may
+define several `Target`s, such as the library, binaries, integration
+test or examples. Targets are crates (each target defines a crate
+root, like `src/lib.rs` or `examples/foo.rs`) and are what is actually
+compiled by `rustc`.
+
+A typical package defines the single library target and several
+auxiliary ones. Packages are a unit of dependency in Cargo, and when
+package `foo` depends on package `bar`, that means that each target
+from `foo` needs the library target from `bar`.
+
+`PackageId` is the unique identifier of a (possibly remote)
+package. It consist of three components: name, version and source
+id. Source is the place where the source code for package comes
+from. Typical sources are crates.io, a git repository or a folder on
+the local hard drive.
+
+`Resolve` is the representation of a direct acyclic graph of package
+dependencies, which uses `PackageId`s for nodes. This is the data
+structure that is saved to the lock file. If there is no lockfile,
+Cargo constructs a resolve by finding a graph of packages which
+matches declared dependency specification according to semver.
+
+
+## Persistence
+
+Cargo is a non-daemon command line application, which means that all
+the information used by Cargo must be persisted on the hard drive. The
+main sources of information are `Cargo.toml` and `Cargo.lock` files,
+`.cargo/config` configuration files and the globally shared registry
+of packages downloaded from crates.io, usually located at
+`~/.cargo/registry`. See `src/sources/registry` for the specifics of
+the registry storage format.
+
+
+## Concurrency
+
+Cargo itself is mostly single hreaded, but it schedules compiler jobs
+in parallel if possible. However there can be several different
+instances of Cargo running concurrently on the system, so Cargo uses
+file locking when accessing potentially shared data like the registry
+or the target directory.
+
+
+## Tests
+
+Cargo has an impressive test suite located in the `tests` folder. Most
+of the test are integration: a project structure with `Cargo.toml` and
+rust source code is created in a temporary directory, `cargo` binary
+is invoked via `std::process::Command` and then stdout and stderr are
+verified against the expected output. To simplify testing, several
+macros of the form `[MACRO]` are used in the expected output. For
+example, `[..]` matches any string and `[/]` matches `/` on Unixes and
+`\` on windows.