diff --git a/doc/cupboard.md b/doc/cupboard.md
new file mode 100644
index 0000000..329c1f0
--- /dev/null
+++ b/doc/cupboard.md
@@ -0,0 +1,275 @@
+# Cupboard
+
+
+
+## Basic Concepts
+
+* **cupboard**: corresponds to the concept of a database. A `cupboard` must be
+  opened in a particular directory on a filesystem. Internally implemented as a
+  Berkeley DB environment. `cupboard` objects contain `shelf` objects.
+* **shelf**: contains objects and indices on those objects. Multiple shelves may
+  be convenient for some applications. Other applications might just store
+  everything on the default shelf.
+* **index**: allows rapid access to an object stored on a shelf by the value of
+  a particular slot in that object. Indices may be `:unique`, in which case
+  insertion of a duplicate results in an error, or `:any`, in which case
+  duplicates are permitted.
+* **transaction**: creates an atomic unit of work to be performed on a
+  `cupboard`. A transaction may affect multiple shelves. Transactions can be
+  committed or rolled back. Berkeley DB transactions support deadlock detection,
+  and, Cupboard transactions will attempt to retry in case of
+  deadlock. Therefore, code wrapped in a transaction should be free of all side
+  effects (except for those affecting the database).
+
+
+
+## API Concepts
+
+In the interest of clarity, all examples here use `cb` as an alias of the
+`cupboard.core` package:
+
+    (ns application.package
+      (:require [cupboard.core :as cb]))
+
+Most functions of the Cupboard API have keyword parameters. Most of these
+keywords are optional.
+
+In particular, when working with the default cupboard, the default shelf, and
+the default transaction, `:cupboard`, `:shelf-name`, and `:txn` keyword
+arguments, respectively, should be omitted. Unless stated otherwise, all
+functions in the Cupboard API support these keywords, and have the following
+meanings:
+
+* `:cupboard`: if omitted, use the dynamic variable `cupboard.core/*cupboard*`
+  as the cupboard for the function.
+* `:shelf-name`: if omitted, use the default shelf name. This default name is
+  given by `cupboard.core/*default-shelf-name*`.
+* `:txn`: if omitted, use the dynamic variable `cupboard.core/*txn*` as the
+  transaction variable for the function.
+
+
+
+## Cupboards
+
+Functions which manipulate cupboards do not support the default `:cupboard`,
+`:shelf-name`, and `:txn` keywords because they do not make sense in this context.
+
+* `(cb/open-cupboard path)` opens a cupboard at the specified path and returns it.
+* `(cb/open-cupboard! path)` opens a cupboard at the specified path, and sets
+  `cupboard.core/*cupboard*` to its value.
+* `(cb/close-cupboard cupboard-var)` closes the cupboard referred to by `cupboard-var`.
+* `(cb/close-cupboard!)` closes the default cupboard, `cupboard.core/*cupboard*`.
+* `(cb/with-open-cupboard [path] body-forms)` makes sure that the
+  body forms execute inside an open default cupboard, and closes it when done. Be
+  careful with spawning threads in side the body, since the scope of the body
+  may end before the threads end their work.
+* `(cb/with-open-cupboard [cupboard-var path] body-forms)` works as above, but binds the
+  open cupboard to `cupboard-var`.
+
+A cupboard may be opened for read-only access using the `:read-only` keyword:
+
+* `(cb/open-cupboard! "/tmp/example" :read-only true)`
+* `(cb/with-open-cupboard [my-cupboard "/tmp/example" :read-only true] ...)`
+
+
+
+## Objects
+
+### Defining
+
+Use the `cb/defpersist` macro to define objects, e.g.:
+
+    (cb/defpersist object-name
+      ((:slot-1 :index :unique)
+       (:slot-2 :index :any)
+       (:slot-3)))
+
+Here, `:slot-1` has a unique index, `:slot-2` has a non-unique index, and
+`:slot-3` has no index at all. Non-indexed slots may not be used for query
+clauses.
+
+Slot names should be Clojure keywords.
+
+`cb/defpersist` expands into Clojure's `defstruct` form, so all objects use
+indexed addressing of structs, rather than hashed addressing of hash-maps.
+
+
+### Instantiating
+
+Use the `cb/make-instance` multimethod to instantiate objects defined using
+`cb/defpersist`, e.g.:
+
+* `(cb/make-instance object-name [value-1 value-2 value-3 ...])`
+
+This form will instantiate the object, write it to the default shelf of the
+default open cupboard, and return it.
+
+* `(cb/make-instance object-name [value-1 value-2 value-3] :save false)` will
+  not write the object. This is typically only useful for testing.
+
+Newly-created objects have metadata which describes how Cupboard persists
+them. Of some interest is the `:primary-key` entry in the object's metadata; it
+is currently implemented as a randomly-generated UUID.
+
+
+### Modifying
+
+If you hold on to a reference of an object created using `cb/make-instance`, or
+retrieved using `cb/retrieve`, you may write it directly. This relies on the
+`:primary-key` metadata entry of the object.
+
+* `(cb/save object)`
+
+Since modifying objects and saving them to the database should be fairly common,
+Cupboard provides helper functions for doing so. `cb/passoc!` adds the given
+key-value pair, or multiple pairs, to the given object, saves, and returns
+it. `cb/pdissoc!` removes keys and saves the object.
+
+* `(cb/passoc! object new-key new-value)`
+* `(cb/passoc! object [new-key-1 new-value-1 new-key-2 new-value-2 ...])`
+* `(cb/pdissoc! object key)`
+* `(cb/pdissoc! object [key-1 key-2 ...])`
+
+An object may be also be removed:
+
+* `(cb/delete object)`
+
+`cb/passoc!`, `cb/pdissoc!`, `cb/delete` are most useful in query callbacks.
+
+
+### Retrieving
+
+Objects in Cupboard may only be retrieved by indexed slot values.
+
+* `(cb/retrieve index-slot indexed-value)`
+
+`cb/retrieve` returns an object directly when retrieved using a unique
+index. Otherwise, it falls back on `cb/query` and returns a lazy sequence of
+objects.
+
+
+
+## Shelves
+
+* `(cb/remove-shelf shelf-name :cupboard cupboard-var)` deletes the given shelf
+  by name.
+* `(cb/list-shelves :cupboard cupboard-var)` returns a list of all shelves on
+  the given cupboard.
+* `(cb/shelf-count :cupboard cupboard-var :shelf-name shelf-name)` counts the
+  number of objects on the given shelf.
+* `(cb/clear-shelf :cupboard cupboard-var :shelf-name shelf-name)` deletes all
+  objects on the given shelf.
+
+
+
+## Transactions
+
+* `(cb/begin-txn)` returns an open transaction. It must be either
+  committed or rolled back.
+* `(cb/commit)` commits an open transaction.
+* `(cb/rollback)` rolls back an open transaction.
+* `(cb/with-txn [options] body-forms)` executes body forms in an open
+  transaction and commits it, unless `(cb/rollback)` is executed somewhere in
+  the body.
+
+Deadlocks are part of life when dealing with Berkeley DB transactions, and
+Cupboard tries to handle them gracefully. In particular, you may specify the
+following options to `cb/with-txn`:
+
+* `:max-attempts` specifies how many times a transaction retries if it runs into
+  a deadlock. Default: 5.
+* `:retry-delay-msec` specifies how long to wait before retrying a deadlocked
+  transaction. Default: 10 msec.
+
+When deadlock occurs inside a `cb/with-txn` form, the transaction rolls back,
+waits out the specified delay, and tries again, up to `:max-attempts` times. If
+this does not work, Cupboard throws an exception.
+
+Since code inside `cb/with-txn` may retry, it must not have any side effects!
+
+
+
+## Queries
+
+Cupboard supports picking out a set of objects by a set of criteria from a
+particular shelf.
+
+* `(cb/query clause*)`
+
+Each clause is an s-expression:
+
+* `(rule-fn indexed-slot value)`
+
+`rule-fn` are functions from the following sets:
+
+* From `clojure.core`: `= < <= > >=`
+* From `cupboard.utils'`: `starts-with date= date> date>= date< date<=`
+
+For example:
+
+    (cb/query (starts-with :name "J")
+              (>= :age 25)
+              (< :age 30))
+
+Two notes about performance:
+
+1. Note that only using `=` clauses, i.e., performing a natural join, yields the
+   best performance, e.g., `(cb/query (= :key1 val1) (= :key2 val2)`.
+2. When using ranges anywhere in the query, try to order the clauses in such a
+   way that the first clause reduces the result set as much as possible. Due to
+   limitations of JE, Cupboard cannot determine the optimal order by
+   itself. (This will hopefully be fixed in a future version of JE.)
+
+If no clauses are specified, `cb/query` returns the entire contents of the
+shelf. Be careful.
+
+`cb/query` has two important keyword parameters.
+
+* `:limit` reduces the number of returned entries to the given number.
+* `:callback` specifies an optional function to be called on each object in the
+  query's result set.
+
+`:callback` allows deleting and updating elements in shelves, provided they meet
+the query's criteria. Examples:
+
+    (cb/query (date< :registered (localdate "2009-01-01"))
+              :callback cb/delete)
+    (cb/query (date>= :registered (localdate "2009-06-01"))
+              :callback #(cb/passoc! % :status :new))
+
+Note that queries using explicit `:cupboard`, `:shelf-name`, and `:txn` values
+must explicitly close over those values in the callback definition:
+
+    (cb/query (date= :registered (localdate "2009-09-01"))
+              :cupboard my-cupboard
+              :shelf-name "users"
+              :callback #(cb/passoc! % :status :sept1
+                                       :cupboard my-cupboard
+                                       :shelf-name "users"))
+
+
+
+## Hints and Warnings
+
+* A shelf may contain multiple object types, but doing so requires care. If
+  their indices have overlapping slot names (e.g., if they all have `:login`
+  slots), then queries may return objects of multiple types.
+* Saving an object with a new index defined on it may take time, since other
+  objects on that shelf will have to be scanned to see if they need to be added
+  to that index.
+* Deleting an index currently requires using `cupboard.bdb.je` functions. This
+  will be fixed in a future version.
+* Cupboard uses Joda Time dates. `cupboard.utils/localdate` returns a simple
+  date without a time zone attached. `cupboard.utils/localtime` returns a simple
+  time. `cupboard.utils/datetime` returns a full timestamp, with millisecond
+  precision and a time zone. `cupboard.utils/localdatetime` returns a full
+  timestamp, with millisecond precision, but no time zone. All accept date and
+  time strings in ISO 8601 format.
+* Do not mix types in indexed values. In other words, an index on slot :id
+  should always contain values of the same type, not a mixture of strings and
+  integers, for example. The order of the values `"123"` and `123` relative to
+  each other is not defined in Cupboard.
+* Do not make indices on slots containing Clojure `hash-map` values. Although
+  Cupboard will save such a slot containing a hash map, the lack of ordering
+  guarantees will lead to unexpected results when attempting to retrieve by that
+  index.