diff --git a/openerp/addons/base/tests/test_api.py b/openerp/addons/base/tests/test_api.py
index 1f5203260e6..7deeafa946a 100644
--- a/openerp/addons/base/tests/test_api.py
+++ b/openerp/addons/base/tests/test_api.py
@@ -31,12 +31,10 @@ class TestAPI(common.TransactionCase):
         self.assertTrue(ids)
         self.assertTrue(partners)
 
-        # partners and its contents are instance of the model, and share its ormcache
+        # partners and its contents are instance of the model
         self.assertIsRecordset(partners, 'res.partner')
-        self.assertIs(partners._ormcache, self.env['res.partner']._ormcache)
         for p in partners:
             self.assertIsRecord(p, 'res.partner')
-            self.assertIs(p._ormcache, self.env['res.partner']._ormcache)
 
         self.assertEqual([p.id for p in partners], ids)
         self.assertEqual(self.env['res.partner'].browse(ids), partners)
diff --git a/openerp/models.py b/openerp/models.py
index 73782dea625..2f820cec4e2 100644
--- a/openerp/models.py
+++ b/openerp/models.py
@@ -827,9 +827,6 @@ class BaseModel(object):
                 "TransientModels must have log_access turned on, " \
                 "in order to implement their access rights policy"
 
-        # prepare ormcache, which must be shared by all instances of the model
-        cls._ormcache = {}
-
     @api.model
     @ormcache()
     def _is_an_ordinary_table(self):
@@ -1827,7 +1824,7 @@ class BaseModel(object):
         ``tools.ormcache`` or ``tools.ormcache_multi``.
         """
         try:
-            self._ormcache.clear()
+            self.pool.cache.clear_prefix((self.pool.db_name, self._name))
             self.pool._any_cache_cleared = True
         except AttributeError:
             pass
diff --git a/openerp/modules/registry.py b/openerp/modules/registry.py
index c6e3e2ed1d2..5c9870a488a 100644
--- a/openerp/modules/registry.py
+++ b/openerp/modules/registry.py
@@ -104,6 +104,10 @@ class Registry(Mapping):
         """ Same as ``self[model_name]``. """
         return self.models[model_name]
 
+    @lazy_property
+    def cache(self):
+        return RegistryManager.cache
+
     @lazy_property
     def pure_function_fields(self):
         """ Return the list of pure function fields (field objects) """
@@ -287,6 +291,7 @@ class RegistryManager(object):
 
     """
     _registries = None
+    _cache = None
     _lock = threading.RLock()
     _saved_lock = None
 
@@ -307,6 +312,18 @@ class RegistryManager(object):
             cls._registries = LRU(size)
         return cls._registries
 
+    @classproperty
+    def cache(cls):
+        """ Return the global LRU ormcache. Its keys are tuples with the
+        following structure: (db_name, model_name, method, args...).
+        """
+        with cls.lock():
+            if cls._cache is None:
+                # we allocate one cache entry per 32KB of memory
+                size = max(8192, int(config['limit_memory_soft'] / 32768))
+                cls._cache = LRU(size)
+            return cls._cache
+
     @classmethod
     def lock(cls):
         """ Return the current registry lock. """
diff --git a/openerp/tools/cache.py b/openerp/tools/cache.py
index 57dea0888c1..ab387de361a 100644
--- a/openerp/tools/cache.py
+++ b/openerp/tools/cache.py
@@ -52,16 +52,11 @@ class ormcache(object):
                 (100*float(self.stat_hit))/(self.stat_miss+self.stat_hit))
 
     def lru(self, model):
-        ormcache = model._ormcache
-        try:
-            d = ormcache[self.method]
-        except KeyError:
-            d = ormcache[self.method] = lru.LRU(self.size)
-        return d
+        return model.pool.cache, (model.pool.db_name, model._name, self.method)
 
     def lookup(self, method, *args, **kwargs):
-        d = self.lru(args[0])
-        key = args[self.skiparg:]
+        d, key0 = self.lru(args[0])
+        key = key0 + args[self.skiparg:]
         try:
             r = d[key]
             self.stat_hit += 1
@@ -76,12 +71,12 @@ class ormcache(object):
 
     def clear(self, model, *args):
         """ Remove *args entry from the cache or all keys if *args is undefined """
-        d = self.lru(model)
+        d, key0 = self.lru(model)
         if args:
             logger.warn("ormcache.clear arguments are deprecated and ignored "
                         "(while clearing caches on (%s).%s)",
                         model._name, self.method.__name__)
-        d.clear()
+        d.clear_prefix(key0)
         model.pool._any_cache_cleared = True
 
 
@@ -97,7 +92,7 @@ class ormcache_context(ormcache):
         return super(ormcache_context, self).__call__(method)
 
     def lookup(self, method, *args, **kwargs):
-        d = self.lru(args[0])
+        d, key0 = self.lru(args[0])
 
         # Note. The decorator() wrapper (used in __call__ above) will resolve
         # arguments, and pass them positionally to lookup(). This is why context
@@ -109,7 +104,7 @@ class ormcache_context(ormcache):
         ckey = [(k, context[k]) for k in self.accepted_keys if k in context]
 
         # Beware: do not take the context from args!
-        key = args[self.skiparg:self.context_pos] + tuple(ckey)
+        key = key0 + args[self.skiparg:self.context_pos] + tuple(ckey)
         try:
             r = d[key]
             self.stat_hit += 1
@@ -130,8 +125,8 @@ class ormcache_multi(ormcache):
         self.multi = multi
 
     def lookup(self, method, *args, **kwargs):
-        d = self.lru(args[0])
-        base_key = args[self.skiparg:self.multi] + args[self.multi+1:]
+        d, key0 = self.lru(args[0])
+        base_key = key0 + args[self.skiparg:self.multi] + args[self.multi+1:]
         ids = args[self.multi]
         result = {}
         missed = []
diff --git a/openerp/tools/lru.py b/openerp/tools/lru.py
index 13b76f387b7..c9de62130ce 100644
--- a/openerp/tools/lru.py
+++ b/openerp/tools/lru.py
@@ -119,4 +119,12 @@ class LRU(object):
         self.first = None
         self.last = None
 
+    @synchronized()
+    def clear_prefix(self, prefix):
+        """ Remove from `self` all the items with the given `prefix`. """
+        n = len(prefix)
+        for key in self.keys():
+            if key[:n] == prefix:
+                del self[key]
+
 # vim:expandtab:smartindent:tabstop=4:softtabstop=4:shiftwidth=4: