diff --git a/.circleci/config.yml b/.circleci/config.yml
index a76a14fcd..f98022599 100644
--- a/.circleci/config.yml
+++ b/.circleci/config.yml
@@ -155,11 +155,11 @@ jobs:
           at: .
       - run: cp -r benchmark/ ~/benchmark_backup/
       - run: cp mix.exs ~/benchmark_backup/
-      - run: docker run -e MIX_ENV=benchmark -v ./:/root/app -v ~/results:/root/results -w /root/app membraneframeworklabs/docker_membrane mix run benchmark/run.exs /root/results/results.res
-      - run: git checkout master
+      - run: docker run -e MIX_ENV=benchmark -v ./:/root/app -v ~/results:/root/results -w /root/app membraneframeworklabs/docker_membrane mix run benchmark/run.exs /root/results/feature_branch_results
+      - run: git checkout -f master
       - run: cp  ~/benchmark_backup/mix.exs ~/app
-      - run: docker run -e MIX_ENV=benchmark -v ./:/root/app -v ~/results:/root/results -v ~/benchmark_backup/benchmark:/root/app/benchmark -w /root/app membraneframeworklabs/docker_membrane mix run benchmark/run.exs /root/results/results_ref.res
-      - run: docker run -e MIX_ENV=benchmark -v ./:/root/app -v ~/results:/root/results -v ~/benchmark_backup/benchmark:/root/app/benchmark -w /root/app membraneframeworklabs/docker_membrane mix run benchmark/compare.exs /root/results/results.res /root/results/results_ref.res
+      - run: docker run -e MIX_ENV=benchmark -v ./:/root/app -v ~/results:/root/results -v ~/benchmark_backup/benchmark:/root/app/benchmark -w /root/app membraneframeworklabs/docker_membrane mix run benchmark/run.exs /root/results/master_results
+      - run: docker run -e MIX_ENV=benchmark -v ./:/root/app -v ~/results:/root/results -v ~/benchmark_backup/benchmark:/root/app/benchmark -w /root/app membraneframeworklabs/docker_membrane mix run benchmark/compare.exs /root/results/feature_branch_results /root/results/master_results
       - run: 
           command: rm ~/results/*
           when: always
diff --git a/benchmark/compare.exs b/benchmark/compare.exs
index f82536f66..104522a5e 100644
--- a/benchmark/compare.exs
+++ b/benchmark/compare.exs
@@ -8,7 +8,7 @@
 defmodule Benchmark.Compare do
   require Logger
 
-  def run(results, ref_results) do
+  def run(results, ref_results, results_name, ref_results_name) do
     if Map.keys(results) != Map.keys(ref_results),
       do: raise("Incompatible performance test result files!")
 
@@ -20,17 +20,17 @@ defmodule Benchmark.Compare do
         Enum.map(Map.keys(test_case_results), fn metric_module ->
           """
           METRIC: #{metric_module}
-          #{inspect(Map.get(test_case_results, metric_module), pretty: true)}
-          vs
-          #{inspect(Map.get(test_case_results_ref, metric_module), pretty: true)}
-
+          1. In #{results_name}:
+          #{inspect(Map.get(test_case_results_ref, metric_module), pretty: true, limit: :infinity)}
+          2. In #{ref_results_name}:
+          #{inspect(Map.get(test_case_results, metric_module), pretty: true, limit: :infinity)}
           """
         end)
         |> Enum.join()
 
       Logger.debug("""
       TEST CASE:
-      #{inspect(test_case, pretty: true)}
+      #{inspect(test_case, pretty: true, limit: :infinity)}
 
       #{results_str}
 
@@ -50,4 +50,6 @@ end
 [results_filename, ref_results_filename] = System.argv() |> Enum.take(2)
 results = File.read!(results_filename) |> :erlang.binary_to_term()
 ref_results = File.read!(ref_results_filename) |> :erlang.binary_to_term()
-Benchmark.Compare.run(results, ref_results)
+results_name = String.split(results_filename, "/") |> Enum.at(-1)
+ref_results_name = String.split(ref_results_filename, "/") |> Enum.at(-1)
+Benchmark.Compare.run(results, ref_results, results_name, ref_results_name)
diff --git a/benchmark/metric/final_memory.ex b/benchmark/metric/final_memory.ex
deleted file mode 100644
index d3b14c052..000000000
--- a/benchmark/metric/final_memory.ex
+++ /dev/null
@@ -1,32 +0,0 @@
-defmodule Benchmark.Metric.FinalMemory do
-  @behaviour Benchmark.Metric
-
-  @allowed_worsening_factor 0.5
-
-  @impl true
-  def assert(final_memory, final_memory_ref, test_case) do
-    if final_memory > final_memory_ref * (1 + @allowed_worsening_factor),
-      do:
-        raise(
-          "The memory performance has got worse! For test case: #{inspect(test_case, pretty: true)}
-        the final memory used to be: #{final_memory_ref} MB and now it is: #{final_memory} MB"
-        )
-
-    :ok
-  end
-
-  @impl true
-  def average(final_memory_samples) do
-    Enum.sum(final_memory_samples) / (length(final_memory_samples) * 1_000_000)
-  end
-
-  @impl true
-  def start_meassurement(_opts \\ nil) do
-    :erlang.memory(:total)
-  end
-
-  @impl true
-  def stop_meassurement(starting_memory) do
-    :erlang.memory(:total) - starting_memory
-  end
-end
diff --git a/benchmark/metric/in_progress_memory.ex b/benchmark/metric/in_progress_memory.ex
index 9d5a36996..71925924a 100644
--- a/benchmark/metric/in_progress_memory.ex
+++ b/benchmark/metric/in_progress_memory.ex
@@ -1,7 +1,7 @@
 defmodule Benchmark.Metric.InProgressMemory do
   @behaviour Benchmark.Metric
 
-  @allowed_worsening_factor 0.5
+  @tolerance_factor 0.5
   @sampling_period 100
 
   @impl true
@@ -9,7 +9,7 @@ defmodule Benchmark.Metric.InProgressMemory do
     cumulative_memory = integrate(memory_samples)
     cumulative_memory_ref = integrate(memory_samples_ref)
 
-    if cumulative_memory > cumulative_memory_ref * (1 + @allowed_worsening_factor),
+    if cumulative_memory > cumulative_memory_ref * (1 + @tolerance_factor),
       do:
         raise(
           "The memory performance has got worse! For test case: #{inspect(test_case, pretty: true)}
diff --git a/benchmark/metric/message_queues_length.ex b/benchmark/metric/message_queues_length.ex
index 911f94e96..61ad10494 100644
--- a/benchmark/metric/message_queues_length.ex
+++ b/benchmark/metric/message_queues_length.ex
@@ -1,7 +1,7 @@
 defmodule Benchmark.Metric.MessageQueuesLength do
   @behaviour Benchmark.Metric
 
-  @allowed_worsening_factor 0.5
+  @tolerance_factor 0.5
   @sampling_period 100
 
   @impl true
@@ -10,7 +10,7 @@ defmodule Benchmark.Metric.MessageQueuesLength do
     cumulative_queues_length_ref = integrate(queues_lengths_ref)
 
     if cumulative_queues_length >
-         cumulative_queues_length_ref * (1 + @allowed_worsening_factor),
+         cumulative_queues_length_ref * (1 + @tolerance_factor),
        do:
          raise(
            "The cumulative queues length has got worse! For test case: #{inspect(test_case, pretty: true)}
diff --git a/benchmark/metric/time.ex b/benchmark/metric/time.ex
index 0517bb53a..211738f49 100644
--- a/benchmark/metric/time.ex
+++ b/benchmark/metric/time.ex
@@ -1,11 +1,11 @@
 defmodule Benchmark.Metric.Time do
   @behaviour Benchmark.Metric
 
-  @allowed_worsening_factor 0.1
+  @tolerance_factor 0.15
 
   @impl true
   def assert(time, time_ref, test_case) do
-    if time > time_ref * (1 + @allowed_worsening_factor),
+    if time > time_ref * (1 + @tolerance_factor),
       do:
         raise(
           "The time performance has got worse! For test case: #{inspect(test_case, pretty: true)} the test
diff --git a/benchmark/run.exs b/benchmark/run.exs
index 5125244fc..162cb76bc 100644
--- a/benchmark/run.exs
+++ b/benchmark/run.exs
@@ -54,7 +54,7 @@ defmodule Benchmark.Run do
   alias Membrane.Pad
   alias Benchmark.Run.BranchedFilter
   alias Benchmark.Run.LinearFilter
-  alias Benchmark.Metric.{FinalMemory, InProgressMemory, MessageQueuesLength, Time}
+  alias Benchmark.Metric.{InProgressMemory, MessageQueuesLength, Time}
 
   require Logger
   require Membrane.RCPipeline
@@ -72,7 +72,7 @@ defmodule Benchmark.Run do
       reductions: 1_000,
       max_random: 1,
       number_of_filters: 100,
-      number_of_buffers: 50000,
+      number_of_buffers: 50_000,
       buffer_size: 1
     ],
     linear: [
@@ -92,7 +92,7 @@ defmodule Benchmark.Run do
     with_branches: [
       struct: [{1, 2}, {1, 2}, {2, 1}, {2, 1}],
       reductions: 100,
-      number_of_buffers: 50_000,
+      number_of_buffers: 500_000,
       buffer_size: 1,
       max_random: 10
     ]
@@ -245,7 +245,6 @@ defmodule Benchmark.Run do
     spec = prepare_pipeline(test_type, params)
 
     time_meassurement = Time.start_meassurement()
-    final_memory_meassurement = FinalMemory.start_meassurement()
 
     {:ok, _supervisor_pid, pipeline_pid} =
       Benchmark.Run.Pipeline.start(
@@ -267,7 +266,6 @@ defmodule Benchmark.Run do
       end
 
     time = Time.stop_meassurement(time_meassurement)
-    final_memory = FinalMemory.stop_meassurement(final_memory_meassurement)
     in_progress_memory = InProgressMemory.stop_meassurement(in_progress_memory_meassurment)
 
     message_queues_length =
@@ -277,7 +275,6 @@ defmodule Benchmark.Run do
 
     %{
       Time => time,
-      FinalMemory => final_memory,
       InProgressMemory => in_progress_memory,
       MessageQueuesLength => message_queues_length
     }