diff --git a/substrate/core/consensus/babe/src/epoch_changes.rs b/substrate/core/consensus/babe/src/epoch_changes.rs
index 311271ae15..09a14d2864 100644
--- a/substrate/core/consensus/babe/src/epoch_changes.rs
+++ b/substrate/core/consensus/babe/src/epoch_changes.rs
@@ -190,23 +190,35 @@ impl<Hash, Number> EpochChanges<Hash, Number> where
 		EpochChanges { inner: ForkTree::new() }
 	}
 
-	/// Prune out finalized epochs, except for the ancestor of the finalized block.
+	/// Prune out finalized epochs, except for the ancestor of the finalized
+	/// block. The given slot should be the slot number at which the finalized
+	/// block was authored.
 	pub fn prune_finalized<D: IsDescendentOfBuilder<Hash>>(
 		&mut self,
 		descendent_of_builder: D,
-		_hash: &Hash,
-		_number: Number,
+		hash: &Hash,
+		number: Number,
+		slot: SlotNumber,
 	) -> Result<(), fork_tree::Error<D::Error>> {
-		let _is_descendent_of = descendent_of_builder
+		let is_descendent_of = descendent_of_builder
 			.build_is_descendent_of(None);
 
-		// TODO:
-		// https://github.com/paritytech/substrate/issues/3651
-		//
+		let predicate = |epoch: &PersistedEpoch| match *epoch {
+			PersistedEpoch::Genesis(_, ref epoch_1) =>
+				slot >= epoch_1.end_slot(),
+			PersistedEpoch::Regular(ref epoch_n) =>
+				slot >= epoch_n.end_slot(),
+		};
+
 		// prune any epochs which could not be _live_ as of the children of the
-		// finalized block.
-		// i.e. re-root the fork tree to the oldest ancestor of (hash, number)
-		// where epoch.end_slot() >= slot(hash)
+		// finalized block, i.e. re-root the fork tree to the oldest ancestor of
+		// (hash, number) where epoch.end_slot() >= finalized_slot
+		self.inner.prune(
+			hash,
+			&number,
+			&is_descendent_of,
+			&predicate,
+		)?;
 
 		Ok(())
 	}
@@ -300,6 +312,12 @@ impl<Hash, Number> EpochChanges<Hash, Number> where
 			Err(e) => Err(e),
 		}
 	}
+
+	/// Return the inner fork tree, useful for testing purposes.
+	#[cfg(test)]
+	pub fn tree(&self) -> &ForkTree<Hash, Number, PersistedEpoch> {
+		&self.inner
+	}
 }
 
 /// Type alias to produce the epoch-changes tree from a block type.
diff --git a/substrate/core/consensus/babe/src/lib.rs b/substrate/core/consensus/babe/src/lib.rs
index d2a16bedb8..b4f376b50d 100644
--- a/substrate/core/consensus/babe/src/lib.rs
+++ b/substrate/core/consensus/babe/src/lib.rs
@@ -254,24 +254,6 @@ pub fn start_babe<B, C, SC, E, I, SO, Error>(BabeParams {
 		&inherent_data_providers,
 	)?;
 
-	let epoch_changes = babe_link.epoch_changes.clone();
-	let pruning_task = client.finality_notification_stream()
-		.for_each(move |notification| {
-			// TODO: supply is-descendent-of and maybe write to disk _now_
-			// as opposed to waiting for the next epoch?
-			let res = epoch_changes.lock().prune_finalized(
-				descendent_query(&*client),
-				&notification.hash,
-				*notification.header.number(),
-			);
-
-			if let Err(e) = res {
-				babe_err!("Could not prune expired epoch changes: {:?}", e);
-			}
-
-			future::ready(())
-		});
-
 	babe_info!("Starting BABE Authorship worker");
 	let slot_worker = slots::start_slot_worker(
 		config.0,
@@ -280,9 +262,9 @@ pub fn start_babe<B, C, SC, E, I, SO, Error>(BabeParams {
 		sync_oracle,
 		inherent_data_providers,
 		babe_link.time_source,
-	).map(|_| ());
+	);
 
-	Ok(future::select(slot_worker, pruning_task).map(|_| Ok::<(), ()>(())).compat())
+	Ok(slot_worker.map(|_| Ok::<(), ()>(())).compat())
 }
 
 struct BabeWorker<B: BlockT, C, E, I, SO> {
@@ -889,6 +871,8 @@ impl<B, E, Block, I, RA, PRA> BlockImport<Block> for BabeBlockImport<B, E, Block
 		// this way we can revert it if there's any error
 		let mut old_epoch_changes = None;
 
+		let info = self.client.info().chain;
+
 		if let Some(next_epoch_descriptor) = next_epoch_digest {
 			let next_epoch = epoch.increment(next_epoch_descriptor);
 
@@ -898,21 +882,48 @@ impl<B, E, Block, I, RA, PRA> BlockImport<Block> for BabeBlockImport<B, E, Block
 				epoch.as_ref().epoch_index, hash, slot_number, epoch.as_ref().start_slot);
 			babe_info!("Next epoch starts at slot {}", next_epoch.as_ref().start_slot);
 
-			// track the epoch change in the fork tree
-			let res = epoch_changes.import(
-				descendent_query(&*self.client),
-				hash,
-				number,
-				*block.header.parent_hash(),
-				next_epoch,
-			);
+			// prune the tree of epochs not part of the finalized chain or
+			// that are not live anymore, and then track the given epoch change
+			// in the tree.
+			// NOTE: it is important that these operations are done in this
+			// order, otherwise if pruning after import the `is_descendent_of`
+			// used by pruning may not know about the block that is being
+			// imported.
+			let prune_and_import = || {
+				let finalized_slot = {
+					let finalized_header = self.client.header(&BlockId::Hash(info.finalized_hash))
+						.map_err(|e| ConsensusError::ClientImport(format!("{:?}", e)))?
+						.expect("best finalized hash was given by client; \
+								finalized headers must exist in db; qed");
 
+					find_pre_digest::<Block>(&finalized_header)
+						.expect("finalized header must be valid; \
+								valid blocks have a pre-digest; qed")
+						.slot_number()
+				};
 
-			if let Err(e) = res {
-				let err = ConsensusError::ClientImport(format!("{:?}", e));
+				epoch_changes.prune_finalized(
+					descendent_query(&*self.client),
+					&info.finalized_hash,
+					info.finalized_number,
+					finalized_slot,
+				).map_err(|e| ConsensusError::ClientImport(format!("{:?}", e)))?;
+
+				epoch_changes.import(
+					descendent_query(&*self.client),
+					hash,
+					number,
+					*block.header.parent_hash(),
+					next_epoch,
+				).map_err(|e| ConsensusError::ClientImport(format!("{:?}", e)))?;
+
+				Ok(())
+			};
+
+			if let Err(e) = prune_and_import() {
 				babe_err!("Failed to launch next epoch: {:?}", e);
 				*epoch_changes = old_epoch_changes.expect("set `Some` above and not taken; qed");
-				return Err(err);
+				return Err(e);
 			}
 
 			crate::aux_schema::write_epoch_changes::<Block, _, _>(
@@ -935,10 +946,7 @@ impl<B, E, Block, I, RA, PRA> BlockImport<Block> for BabeBlockImport<B, E, Block
 		// more primary blocks), if there's a tie we go with the longest
 		// chain.
 		block.fork_choice = {
-			let (last_best, last_best_number) = {
-				let info = self.client.info().chain;
-				(info.best_hash, info.best_number)
-			};
+			let (last_best, last_best_number) = (info.best_hash, info.best_number);
 
 			let last_best_weight = if &last_best == block.header.parent_hash() {
 				// the parent=genesis case is already covered for loading parent weight,
diff --git a/substrate/core/consensus/babe/src/tests.rs b/substrate/core/consensus/babe/src/tests.rs
index adfe0f03af..f06cc6f46c 100644
--- a/substrate/core/consensus/babe/src/tests.rs
+++ b/substrate/core/consensus/babe/src/tests.rs
@@ -103,6 +103,7 @@ impl DummyProposer {
 			&BlockId::Hash(self.parent_hash),
 			pre_digests,
 		).unwrap();
+
 		let mut block = match block_builder.bake().map_err(|e| e.into()) {
 			Ok(b) => b,
 			Err(e) => return future::ready(Err(e)),
@@ -597,3 +598,163 @@ fn importing_block_one_sets_genesis_epoch() {
 	).unwrap().unwrap().into_inner();
 	assert_eq!(epoch_for_second_block, genesis_epoch);
 }
+
+#[test]
+fn importing_epoch_change_block_prunes_tree() {
+	use client::backend::Finalizer;
+
+	let mut net = BabeTestNet::new(1);
+
+	let peer = net.peer(0);
+	let data = peer.data.as_ref().expect("babe link set up during initialization");
+
+	let client = peer.client().as_full().expect("Only full clients are used in tests").clone();
+	let mut block_import = data.block_import.lock().take().expect("import set up during init");
+	let epoch_changes = data.link.epoch_changes.clone();
+
+	// This is just boilerplate code for proposing and importing a valid BABE
+	// block that's built on top of the given parent. The proposer takes care
+	// of producing epoch change digests according to the epoch duration (which
+	// is set to 6 slots in the test runtime).
+	let mut propose_and_import_block = |parent_header| {
+		let mut environ = DummyFactory {
+			client: client.clone(),
+			config: data.link.config.clone(),
+			epoch_changes: data.link.epoch_changes.clone(),
+			mutator: Arc::new(|_, _| ()),
+		};
+
+		let mut proposer = environ.init(&parent_header).unwrap();
+		let parent_pre_digest = find_pre_digest::<Block>(&parent_header).unwrap();
+
+		let pre_digest = sr_primitives::generic::Digest {
+			logs: vec![
+				Item::babe_pre_digest(
+					BabePreDigest::Secondary {
+						authority_index: 0,
+						slot_number: parent_pre_digest.slot_number() + 1,
+					},
+				),
+			],
+		};
+
+		let mut block = futures::executor::block_on(proposer.propose_with(pre_digest)).unwrap();
+
+		let seal = {
+			// sign the pre-sealed hash of the block and then
+			// add it to a digest item.
+			let pair = AuthorityPair::from_seed(&[1; 32]);
+			let pre_hash = block.header.hash();
+			let signature = pair.sign(pre_hash.as_ref());
+			Item::babe_seal(signature)
+		};
+
+		let post_hash = {
+			block.header.digest_mut().push(seal.clone());
+			let h = block.header.hash();
+			block.header.digest_mut().pop();
+			h
+		};
+
+		let next_epoch_digest =
+			find_next_epoch_digest::<Block>(&block.header).unwrap();
+
+		let import_result = block_import.import_block(
+			BlockImportParams {
+				origin: BlockOrigin::Own,
+				header: block.header,
+				justification: None,
+				post_digests: vec![seal],
+				body: Some(block.extrinsics),
+				finalized: false,
+				auxiliary: Vec::new(),
+				fork_choice: ForkChoiceStrategy::LongestChain,
+			},
+			Default::default(),
+		).unwrap();
+
+		match import_result {
+			ImportResult::Imported(_) => {},
+			_ => panic!("expected block to be imported"),
+		}
+
+		(post_hash, next_epoch_digest)
+	};
+
+	let mut propose_and_import_blocks = |parent_id, n| {
+		let mut hashes = Vec::new();
+		let mut parent_header = client.header(&parent_id).unwrap().unwrap();
+
+		for _ in 0..n {
+			let (block_hash, _) = propose_and_import_block(parent_header);
+			hashes.push(block_hash);
+			parent_header = client.header(&BlockId::Hash(block_hash)).unwrap().unwrap();
+		}
+
+		hashes
+	};
+
+	// This is the block tree that we're going to use in this test. Each node
+	// represents an epoch change block, the epoch duration is 6 slots.
+	//
+	//    *---- F (#7)
+	//   /                 *------ G (#19) - H (#25)
+	//  /                 /
+	// A (#1) - B (#7) - C (#13) - D (#19) - E (#25)
+	//                              \
+	//                               *------ I (#25)
+
+	// Create and import the canon chain and keep track of fork blocks (A, C, D)
+	// from the diagram above.
+	let canon_hashes = propose_and_import_blocks(BlockId::Number(0), 30);
+
+	// Create the forks
+	let fork_1 = propose_and_import_blocks(BlockId::Hash(canon_hashes[0]), 10);
+	let fork_2 = propose_and_import_blocks(BlockId::Hash(canon_hashes[12]), 15);
+	let fork_3 = propose_and_import_blocks(BlockId::Hash(canon_hashes[18]), 10);
+
+	// We should be tracking a total of 9 epochs in the fork tree
+	assert_eq!(
+		epoch_changes.lock().tree().iter().count(),
+		9,
+	);
+
+	// And only one root
+	assert_eq!(
+		epoch_changes.lock().tree().roots().count(),
+		1,
+	);
+
+	// We finalize block #13 from the canon chain, so on the next epoch
+	// change the tree should be pruned, to not contain F (#7).
+	client.finalize_block(BlockId::Hash(canon_hashes[12]), None, false).unwrap();
+	propose_and_import_blocks(BlockId::Hash(client.info().chain.best_hash), 7);
+
+	// at this point no hashes from the first fork must exist on the tree
+	assert!(
+		!epoch_changes.lock().tree().iter().map(|(h, _, _)| h).any(|h| fork_1.contains(h)),
+	);
+
+	// but the epoch changes from the other forks must still exist
+	assert!(
+		epoch_changes.lock().tree().iter().map(|(h, _, _)| h).any(|h| fork_2.contains(h))
+	);
+
+	assert!(
+		epoch_changes.lock().tree().iter().map(|(h, _, _)| h).any(|h| fork_3.contains(h)),
+	);
+
+	// finalizing block #25 from the canon chain should prune out the second fork
+	client.finalize_block(BlockId::Hash(canon_hashes[24]), None, false).unwrap();
+	propose_and_import_blocks(BlockId::Hash(client.info().chain.best_hash), 8);
+
+	// at this point no hashes from the second fork must exist on the tree
+	assert!(
+		!epoch_changes.lock().tree().iter().map(|(h, _, _)| h).any(|h| fork_2.contains(h)),
+	);
+
+	// while epoch changes from the last fork should still exist
+	assert!(
+		epoch_changes.lock().tree().iter().map(|(h, _, _)| h).any(|h| fork_3.contains(h)),
+	);
+}
diff --git a/substrate/core/utils/fork-tree/src/lib.rs b/substrate/core/utils/fork-tree/src/lib.rs
index 4299918755..b0258a9ca2 100644
--- a/substrate/core/utils/fork-tree/src/lib.rs
+++ b/substrate/core/utils/fork-tree/src/lib.rs
@@ -86,55 +86,45 @@ impl<H, N, V> ForkTree<H, N, V> where
 	N: Ord + Clone,
 	V: Clone,
 {
-	/// Prune all nodes that are not descendents of `hash` according to
-	/// `is_descendent_of`. The given function `is_descendent_of` should return
-	/// `true` if the second hash (target) is a descendent of the first hash
-	/// (base). After pruning the tree it should have one or zero roots. The
-	/// number and order of calls to `is_descendent_of` is unspecified and
-	/// subject to change.
-	pub fn prune<F, E>(
+	/// Prune the tree, removing all non-canonical nodes. We find the node in the
+	/// tree that is the deepest ancestor of the given hash and that passes the
+	/// given predicate. If such a node exists, we re-root the tree to this
+	/// node. Otherwise the tree remains unchanged. The given function
+	/// `is_descendent_of` should return `true` if the second hash (target) is a
+	/// descendent of the first hash (base).
+	pub fn prune<F, E, P>(
 		&mut self,
 		hash: &H,
-		number: N,
-		is_descendent_of: &F
+		number: &N,
+		is_descendent_of: &F,
+		predicate: &P,
 	) -> Result<(), Error<E>>
 		where E: std::error::Error,
-			  F: Fn(&H, &H) -> Result<bool, E>
+			  F: Fn(&H, &H) -> Result<bool, E>,
+			  P: Fn(&V) -> bool,
 	{
-		let mut new_root = None;
-		for node in self.node_iter() {
-			// if the node has a lower number than the one being finalized then
-			// we only keep if it has no children and the finalized block is a
-			// descendent of this node
-			if node.number < number  {
-				if !node.children.is_empty() || !is_descendent_of(&node.hash, hash)? {
-					continue;
-				}
-			}
-
-			// if the node has the same number as the finalized block then it
-			// must have the same hash
-			if node.number == number && node.hash != *hash {
-				continue;
-			}
-
-			// if the node has a higher number then we keep it if it is a
-			// descendent of the finalized block
-			if node.number > number && !is_descendent_of(hash, &node.hash)? {
-				continue;
-			}
-
-			new_root = Some(node);
-			break;
-		}
+		let new_root = self.find_node_where(
+			hash,
+			number,
+			is_descendent_of,
+			predicate,
+		)?;
 
 		if let Some(root) = new_root {
-			self.roots = vec![root.clone()];
+			let mut root = root.clone();
+
+			// we found the deepest ancestor of the finalized block, so we prune
+			// out any children that don't include the finalized block.
+			root.children.retain(|node| {
+				node.number == *number && node.hash == *hash ||
+					node.number < *number && is_descendent_of(&node.hash, hash).unwrap_or(false)
+			});
+
+			self.roots = vec![root];
 		}
 
 		Ok(())
 	}
-
 }
 
 impl<H, N, V> ForkTree<H, N, V> where
@@ -1203,18 +1193,36 @@ mod test {
 
 		tree.prune(
 			&"C",
-			3,
+			&3,
 			&is_descendent_of,
+			&|_| true,
 		).unwrap();
 
 		assert_eq!(
 			tree.roots.iter().map(|node| node.hash).collect::<Vec<_>>(),
-			vec!["C"],
+			vec!["B"],
 		);
 
 		assert_eq!(
 			tree.iter().map(|(hash, _, _)| *hash).collect::<Vec<_>>(),
-			vec!["C", "D", "E"],
+			vec!["B", "C", "D", "E"],
+		);
+
+		tree.prune(
+			&"E",
+			&5,
+			&is_descendent_of,
+			&|_| true,
+		).unwrap();
+
+		assert_eq!(
+			tree.roots.iter().map(|node| node.hash).collect::<Vec<_>>(),
+			vec!["D"],
+		);
+
+		assert_eq!(
+			tree.iter().map(|(hash, _, _)| *hash).collect::<Vec<_>>(),
+			vec!["D", "E"],
 		);
 	}