diff --git a/README.md b/README.md
index eae38fd..d00a182 100644
--- a/README.md
+++ b/README.md
@@ -13,16 +13,42 @@ A brief summary:
 - no SQL
 - objects are serialized (currently in JSON)
 - indexes (simple symlinks on the FS) can be created to improve significantly searches in the db
+- db is fully integrated in the language (basically a simple array with a few more functions)
+
+Also, data can be `cached`.
+The entire base will be kept in memory (if you can), enabling incredible speeds.
 
 ## Limitations
 
-**TODO**: speed tests, elaborate on the matter.
+DODB doesn't fully handle ACID properties:
 
-DODB is not compatible with projects:
-- having an absolute priority on speed,
-  however, DODB is efficient in most cases with the right indexes.
-- having relational data
+- no *atomicity*, you can't chain instructions and rollback if one of them fails ;
+- no *consistency*, there is currently no mechanism to prevent adding invalid values ;
 
+*Isolation* is partially taken into account, with a locking mechanism preventing a few race conditions.
+FYI, in my projects the database is fast enough so I don't even need parallelism (like, by far).
+
+*Durability* is taken into account.
+Data is written on-disk each time it changes.
+
+**NOTE**: what I need is mostly there.
+What DODB doesn't provide, I hack it in a few lines in my app.
+DODB will provide some form of atomicity and consistency at some point, but nothing fancy nor too advanced.
+The whole point is to keep it simple.
+
+## Speed
+
+Since DODB doesn't use SQL and doesn't even try to handle stuff like atomicity or consistency, speed is great.
+Reading data from disk takes about a few dozen microseconds, and not much more when searching an indexed data.
+
+**On my more-than-decade-old, slow-as-fuck machine**, the simplest possible SQL request to Postgres takes about 100 to 900 microseconds.
+To reach a data on-disk: 13 microseconds.
+To reach a data that is indexed: same thing, 13 microseconds, since it's just a link.
+
+With the `cached` version of DODB, there is not even deserialization happening, so 7 nanoseconds.
+
+**NOTE:** of course SQL and DODB cannot be fairly compared based on performance since they don't have the same properties.
+But still, this is the kind of speed you can get with the tool.
 
 # Installation