diff --git a/p2p/host/peerstore/pstoreds/addr_book.go b/p2p/host/peerstore/pstoreds/addr_book.go
index fe56b185af..7e95a203e4 100644
--- a/p2p/host/peerstore/pstoreds/addr_book.go
+++ b/p2p/host/peerstore/pstoreds/addr_book.go
@@ -282,7 +282,21 @@ func (ab *dsAddrBook) AddAddrs(p peer.ID, addrs []ma.Multiaddr, ttl time.Duratio
 
 // ConsumePeerRecord adds addresses from a signed peer.PeerRecord (contained in
 // a record.Envelope), which will expire after the given TTL.
-// See https://godoc.org/github.com/libp2p/go-libp2p/core/peerstore#CertifiedAddrBook for more details.
+// See https://godoc.org/github.com/libp2p/go-libp2p/core/peerstore#CertifiedAddrBook
+// for more details.
+//
+// The signed peer record's Seq is treated as monotonic per peer: a record
+// with a Seq lower than the last accepted one is rejected. Equal Seq is
+// accepted as a TTL refresh.
+//
+// When a newer signed record is accepted, addrs that were present in the
+// previously stored signed record but absent in the new one are evicted, so
+// the peerstore reflects the peer's current self-advertised set instead of
+// the union of every record we have ever seen. Unsigned addrs (added via
+// AddAddr / SetAddr from sources like DHT gossip, or from an identify
+// exchange where the peer did not send a signed record) are not touched, and
+// addrs held by a live connection (TTL >= ConnectedAddrTTL) are also kept so
+// active sessions are not dropped.
 func (ab *dsAddrBook) ConsumePeerRecord(recordEnvelope *record.Envelope, ttl time.Duration) (bool, error) {
 	r, err := recordEnvelope.Record()
 	if err != nil {
@@ -303,6 +317,15 @@ func (ab *dsAddrBook) ConsumePeerRecord(recordEnvelope *record.Envelope, ttl tim
 	}
 
 	addrs := cleanAddrs(rec.Addrs, rec.PeerID)
+
+	// Diff against the previously stored signed record so we can drop addrs
+	// the peer no longer advertises before adding the new ones.
+	if superseded := ab.supersededSignedAddrs(rec.PeerID, addrs); len(superseded) > 0 {
+		if err := ab.deleteAddrs(rec.PeerID, superseded); err != nil {
+			return false, err
+		}
+	}
+
 	err = ab.setAddrs(rec.PeerID, addrs, ttl, ttlExtend, true)
 	if err != nil {
 		return false, err
@@ -315,6 +338,62 @@ func (ab *dsAddrBook) ConsumePeerRecord(recordEnvelope *record.Envelope, ttl tim
 	return true, nil
 }
 
+// supersededSignedAddrs returns addrs that were present in the previously
+// stored signed peer record for p but are absent in newAddrs. Addrs held by
+// a live connection (TTL >= ConnectedAddrTTL) are excluded so an active
+// session is not torn down when the peer rotates its advertised set.
+func (ab *dsAddrBook) supersededSignedAddrs(p peer.ID, newAddrs []ma.Multiaddr) []ma.Multiaddr {
+	prevEnv := ab.GetPeerRecord(p)
+	if prevEnv == nil {
+		return nil
+	}
+	prev, err := prevEnv.Record()
+	if err != nil {
+		return nil
+	}
+	prevRec, ok := prev.(*peer.PeerRecord)
+	if !ok {
+		return nil
+	}
+
+	newSet := make(map[string]struct{}, len(newAddrs))
+	for _, a := range newAddrs {
+		newSet[string(a.Bytes())] = struct{}{}
+	}
+
+	pr, err := ab.loadRecord(p, true, false)
+	if err != nil {
+		return nil
+	}
+	pr.RLock()
+	connected := make(map[string]struct{})
+	for _, a := range pr.Addrs {
+		if ttlIsConnected(time.Duration(a.Ttl)) {
+			connected[string(a.Addr)] = struct{}{}
+		}
+	}
+	pr.RUnlock()
+
+	superseded := make([]ma.Multiaddr, 0, len(prevRec.Addrs))
+	for _, a := range prevRec.Addrs {
+		key := string(a.Bytes())
+		if _, still := newSet[key]; still {
+			continue
+		}
+		if _, isConn := connected[key]; isConn {
+			continue
+		}
+		superseded = append(superseded, a)
+	}
+	return superseded
+}
+
+// ttlIsConnected reports whether the given TTL marks the address as held by
+// a live connection.
+func ttlIsConnected(ttl time.Duration) bool {
+	return ttl >= pstore.ConnectedAddrTTL
+}
+
 func (ab *dsAddrBook) latestPeerRecordSeq(p peer.ID) uint64 {
 	pr, err := ab.loadRecord(p, true, false)
 	if err != nil {
diff --git a/p2p/host/peerstore/pstoreds/ds_test.go b/p2p/host/peerstore/pstoreds/ds_test.go
index a240de3a1c..c5aeb70beb 100644
--- a/p2p/host/peerstore/pstoreds/ds_test.go
+++ b/p2p/host/peerstore/pstoreds/ds_test.go
@@ -5,12 +5,17 @@ import (
 	"testing"
 	"time"
 
+	"github.com/libp2p/go-libp2p/core/crypto"
+	"github.com/libp2p/go-libp2p/core/peer"
 	pstore "github.com/libp2p/go-libp2p/core/peerstore"
+	"github.com/libp2p/go-libp2p/core/record"
+	"github.com/libp2p/go-libp2p/core/test"
 	pt "github.com/libp2p/go-libp2p/p2p/host/peerstore/test"
 
 	mockclock "github.com/benbjohnson/clock"
 	ds "github.com/ipfs/go-datastore"
 	"github.com/ipfs/go-datastore/sync"
+	ma "github.com/multiformats/go-multiaddr"
 	"github.com/stretchr/testify/require"
 )
 
@@ -72,6 +77,60 @@ func TestDsAddrBook(t *testing.T) {
 	}
 }
 
+// TestDsConsumePeerRecordReplacesStaleAddrs verifies replace-semantics on a
+// newer signed peer record: addrs dropped from the new record are evicted,
+// while unsigned addrs and addrs held by a live connection are kept.
+func TestDsConsumePeerRecordReplacesStaleAddrs(t *testing.T) {
+	for name, dsFactory := range dstores {
+		t.Run(name, func(t *testing.T) {
+			opts := DefaultOpts()
+			store, closeDs := dsFactory(t)
+			defer closeDs()
+			ab, err := NewAddrBook(context.Background(), store, opts)
+			require.NoError(t, err)
+			defer ab.Close()
+
+			priv, _, err := test.RandTestKeyPair(crypto.Ed25519, 256)
+			require.NoError(t, err)
+			id, err := peer.IDFromPrivateKey(priv)
+			require.NoError(t, err)
+
+			keep := ma.StringCast("/ip4/1.2.3.4/tcp/1")
+			drop := ma.StringCast("/ip4/1.2.3.4/tcp/2")
+			unsigned := ma.StringCast("/ip4/1.2.3.4/tcp/3")
+			connected := ma.StringCast("/ip4/1.2.3.4/tcp/4")
+
+			rec1 := peer.NewPeerRecord()
+			rec1.PeerID = id
+			rec1.Seq = 1
+			rec1.Addrs = []ma.Multiaddr{keep, drop, connected}
+			env1, err := record.Seal(rec1, priv)
+			require.NoError(t, err)
+
+			accepted, err := ab.ConsumePeerRecord(env1, time.Hour)
+			require.NoError(t, err)
+			require.True(t, accepted)
+
+			ab.AddAddr(id, connected, pstore.ConnectedAddrTTL)
+			ab.AddAddr(id, unsigned, time.Hour)
+			require.ElementsMatch(t, []ma.Multiaddr{keep, drop, connected, unsigned}, ab.Addrs(id))
+
+			rec2 := peer.NewPeerRecord()
+			rec2.PeerID = id
+			rec2.Seq = 2
+			rec2.Addrs = []ma.Multiaddr{keep}
+			env2, err := record.Seal(rec2, priv)
+			require.NoError(t, err)
+
+			accepted, err = ab.ConsumePeerRecord(env2, time.Hour)
+			require.NoError(t, err)
+			require.True(t, accepted)
+
+			require.ElementsMatch(t, []ma.Multiaddr{keep, connected, unsigned}, ab.Addrs(id))
+		})
+	}
+}
+
 func TestDsKeyBook(t *testing.T) {
 	for name, dsFactory := range dstores {
 		t.Run(name, func(t *testing.T) {
diff --git a/p2p/host/peerstore/pstoremem/addr_book.go b/p2p/host/peerstore/pstoremem/addr_book.go
index 4c389cfcd1..a09c172312 100644
--- a/p2p/host/peerstore/pstoremem/addr_book.go
+++ b/p2p/host/peerstore/pstoremem/addr_book.go
@@ -44,7 +44,9 @@ func ttlIsConnected(ttl time.Duration) bool {
 
 type peerRecordState struct {
 	Envelope *record.Envelope
-	Seq      uint64
+	// Seq is the sequence number from the stored signed peer record. Newer
+	// records (higher Seq) supersede older ones for the same peer.
+	Seq uint64
 }
 
 // Essentially Go stdlib's Priority Queue example
@@ -289,8 +291,23 @@ func (mab *memoryAddrBook) AddAddrs(p peer.ID, addrs []ma.Multiaddr, ttl time.Du
 	mab.addAddrs(p, addrs, ttl)
 }
 
-// ConsumePeerRecord adds addresses from a signed peer.PeerRecord, which will expire after the given TTL.
-// See https://godoc.org/github.com/libp2p/go-libp2p/core/peerstore#CertifiedAddrBook for more details.
+// ConsumePeerRecord adds addresses from a signed peer.PeerRecord, which will
+// expire after the given TTL. See
+// https://godoc.org/github.com/libp2p/go-libp2p/core/peerstore#CertifiedAddrBook
+// for more details.
+//
+// The signed peer record's Seq is treated as monotonic per peer: a record with
+// a Seq lower than the last accepted one is rejected. Equal Seq is accepted as
+// a TTL refresh.
+//
+// When a newer signed record is accepted, addrs that were present in the
+// previously stored signed record but absent in the new one are evicted, so
+// the peerstore reflects the peer's current self-advertised set instead of
+// the union of every record we have ever seen. Unsigned addrs (added via
+// AddAddr / SetAddr from sources like DHT gossip, or from an identify
+// exchange where the peer did not send a signed record) are not touched, and
+// addrs held by a live connection (TTL >= ConnectedAddrTTL) are also kept so
+// active sessions are not dropped.
 func (mab *memoryAddrBook) ConsumePeerRecord(recordEnvelope *record.Envelope, ttl time.Duration) (bool, error) {
 	r, err := recordEnvelope.Record()
 	if err != nil {
@@ -316,6 +333,33 @@ func (mab *memoryAddrBook) ConsumePeerRecord(recordEnvelope *record.Envelope, tt
 	if !found && len(mab.signedPeerRecords) >= mab.maxSignedPeerRecords {
 		return false, errors.New("too many signed peer records")
 	}
+
+	// Drop addrs from the previous signed record that are absent in the
+	// new one; addrs held by a live connection are preserved so we don't
+	// drop an active session if the peer rotates its advertised set. The
+	// prior addr set is recovered by decoding the stored envelope; that
+	// call caches on first access (core/record/envelope.go), so repeated
+	// lookups are cheap.
+	if found {
+		if prevRec := prevSignedAddrs(lastState); len(prevRec) > 0 {
+			newAddrSet := make(map[string]struct{}, len(rec.Addrs))
+			for _, a := range rec.Addrs {
+				newAddrSet[string(a.Bytes())] = struct{}{}
+			}
+			for _, a := range prevRec {
+				key := string(a.Bytes())
+				if _, still := newAddrSet[key]; still {
+					continue
+				}
+				ea, ok := mab.addrs.Addrs[rec.PeerID][key]
+				if !ok || ea.IsConnected() {
+					continue
+				}
+				mab.addrs.Delete(ea)
+			}
+		}
+	}
+
 	mab.signedPeerRecords[rec.PeerID] = &peerRecordState{
 		Envelope: recordEnvelope,
 		Seq:      rec.Seq,
@@ -324,6 +368,24 @@ func (mab *memoryAddrBook) ConsumePeerRecord(recordEnvelope *record.Envelope, tt
 	return true, nil
 }
 
+// prevSignedAddrs returns the addrs from the stored signed peer record, or
+// nil if the envelope is absent or can't be decoded. Envelope.Record() caches
+// its result, so repeated calls are cheap.
+func prevSignedAddrs(s *peerRecordState) []ma.Multiaddr {
+	if s == nil || s.Envelope == nil {
+		return nil
+	}
+	r, err := s.Envelope.Record()
+	if err != nil {
+		return nil
+	}
+	pr, ok := r.(*peer.PeerRecord)
+	if !ok {
+		return nil
+	}
+	return pr.Addrs
+}
+
 func (mab *memoryAddrBook) maybeDeleteSignedPeerRecordUnlocked(p peer.ID) {
 	if len(mab.addrs.Addrs[p]) == 0 {
 		delete(mab.signedPeerRecords, p)
diff --git a/p2p/host/peerstore/pstoremem/addr_book_test.go b/p2p/host/peerstore/pstoremem/addr_book_test.go
index 21fa8d751b..72cc474498 100644
--- a/p2p/host/peerstore/pstoremem/addr_book_test.go
+++ b/p2p/host/peerstore/pstoremem/addr_book_test.go
@@ -8,7 +8,11 @@ import (
 	"testing"
 	"time"
 
+	"github.com/libp2p/go-libp2p/core/crypto"
 	"github.com/libp2p/go-libp2p/core/peer"
+	"github.com/libp2p/go-libp2p/core/peerstore"
+	"github.com/libp2p/go-libp2p/core/record"
+	"github.com/libp2p/go-libp2p/core/test"
 	ma "github.com/multiformats/go-multiaddr"
 	"github.com/stretchr/testify/require"
 )
@@ -186,6 +190,56 @@ func TestPeerLimits(t *testing.T) {
 	require.Equal(t, 1024, ab.addrs.NumUnconnectedAddrs())
 }
 
+// TestConsumePeerRecordReplacesStaleAddrs verifies replace-semantics on a
+// newer signed peer record: addrs dropped from the new record are evicted,
+// while unsigned addrs and addrs held by a live connection are kept.
+func TestConsumePeerRecordReplacesStaleAddrs(t *testing.T) {
+	ab := NewAddrBook()
+	defer ab.Close()
+
+	priv, _, err := test.RandTestKeyPair(crypto.Ed25519, 256)
+	require.NoError(t, err)
+	id, err := peer.IDFromPrivateKey(priv)
+	require.NoError(t, err)
+
+	keep := ma.StringCast("/ip4/1.2.3.4/tcp/1")
+	drop := ma.StringCast("/ip4/1.2.3.4/tcp/2")
+	unsigned := ma.StringCast("/ip4/1.2.3.4/tcp/3")
+	connected := ma.StringCast("/ip4/1.2.3.4/tcp/4")
+
+	rec1 := peer.NewPeerRecord()
+	rec1.PeerID = id
+	rec1.Seq = 1
+	rec1.Addrs = []ma.Multiaddr{keep, drop, connected}
+	env1, err := record.Seal(rec1, priv)
+	require.NoError(t, err)
+
+	accepted, err := ab.ConsumePeerRecord(env1, time.Hour)
+	require.NoError(t, err)
+	require.True(t, accepted)
+
+	// Pin `connected` via ConnectedAddrTTL and add an unsigned addr.
+	ab.AddAddr(id, connected, peerstore.ConnectedAddrTTL)
+	ab.AddAddr(id, unsigned, time.Hour)
+	require.ElementsMatch(t, []ma.Multiaddr{keep, drop, connected, unsigned}, ab.Addrs(id))
+
+	// Newer record drops `drop` and only mentions `keep`. `drop` must go;
+	// `unsigned` (never in a signed record) and `connected` (held by a
+	// live connection) must stay.
+	rec2 := peer.NewPeerRecord()
+	rec2.PeerID = id
+	rec2.Seq = 2
+	rec2.Addrs = []ma.Multiaddr{keep}
+	env2, err := record.Seal(rec2, priv)
+	require.NoError(t, err)
+
+	accepted, err = ab.ConsumePeerRecord(env2, time.Hour)
+	require.NoError(t, err)
+	require.True(t, accepted)
+
+	require.ElementsMatch(t, []ma.Multiaddr{keep, connected, unsigned}, ab.Addrs(id))
+}
+
 func BenchmarkPeerAddrs(b *testing.B) {
 	sizes := [...]int{1, 10, 100, 1000, 10_000, 100_000, 1000_000}
 	for _, sz := range sizes {
diff --git a/p2p/host/peerstore/test/addr_book_suite.go b/p2p/host/peerstore/test/addr_book_suite.go
index dbe48f5e84..099d64a2fd 100644
--- a/p2p/host/peerstore/test/addr_book_suite.go
+++ b/p2p/host/peerstore/test/addr_book_suite.go
@@ -475,8 +475,8 @@ func testCertifiedAddresses(m pstore.AddrBook, clk *mockClock.Mock) func(*testin
 			t.Error("unable to retrieve signed routing record from addrbook")
 		}
 
-		// Adding a new envelope should clear existing certified addresses.
-		// Only the newly-added ones should remain
+		// A newer signed record drops addrs the peer no longer advertises.
+		// Unsigned addrs (added via plain AddAddrs) are retained.
 		certifiedAddrs = certifiedAddrs[:3]
 		rec4 := peer.NewPeerRecord()
 		rec4.PeerID = id
@@ -488,8 +488,9 @@ func testCertifiedAddresses(m pstore.AddrBook, clk *mockClock.Mock) func(*testin
 		if !accepted {
 			t.Error("expected peer record to be accepted")
 		}
-		// AssertAddressesEqual(t, certifiedAddrs, m.Addrs(id))
-		AssertAddressesEqual(t, allAddrs, m.Addrs(id))
+		expectedAfterRec4 := append([]multiaddr.Multiaddr{}, certifiedAddrs...)
+		expectedAfterRec4 = append(expectedAfterRec4, uncertifiedAddrs...)
+		AssertAddressesEqual(t, expectedAfterRec4, m.Addrs(id))
 
 		// update TTL on signed addrs to -1 to remove them.
 		// the signed routing record should be deleted