diff --git a/management/internals/modules/peers/ephemeral/manager/ephemeral.go b/management/internals/modules/peers/ephemeral/manager/ephemeral.go
index 0f902ea70..8b0a22efc 100644
--- a/management/internals/modules/peers/ephemeral/manager/ephemeral.go
+++ b/management/internals/modules/peers/ephemeral/manager/ephemeral.go
@@ -2,6 +2,7 @@ package manager
 
 import (
 	"context"
+	"math/rand"
 	"sync"
 	"time"
 
@@ -11,44 +12,76 @@ import (
 	"github.com/netbirdio/netbird/management/internals/modules/peers/ephemeral"
 	"github.com/netbirdio/netbird/management/server/activity"
 	nbpeer "github.com/netbirdio/netbird/management/server/peer"
-	"github.com/netbirdio/netbird/management/server/telemetry"
-
 	"github.com/netbirdio/netbird/management/server/store"
+	"github.com/netbirdio/netbird/management/server/telemetry"
 )
 
 const (
-	// cleanupWindow is the time window to wait after nearest peer deadline to start the cleanup procedure.
+	// cleanupWindow is the small grace period added on top of the
+	// staleness horizon before a sweep fires. It absorbs minor clock
+	// skew between the management server and the database and avoids
+	// firing a sweep right at the boundary where last_seen could still
+	// be one tick under the threshold.
 	cleanupWindow = 1 * time.Minute
+
+	// initialLoadMinDelay and initialLoadMaxDelay bracket the random
+	// delay applied before the post-restart catch-up query runs. Spread
+	// across replicas this prevents a thundering herd of catch-up
+	// queries hitting the database simultaneously after a deploy.
+	initialLoadMinDelay = 8 * time.Minute
+	initialLoadMaxDelay = 10 * time.Minute
 )
 
 var (
 	timeNow = time.Now
 )
 
-type ephemeralPeer struct {
-	id        string
-	accountID string
-	deadline  time.Time
-	next      *ephemeralPeer
+// accountEntry is the per-account state held by the cleanup tracker.
+// We don't track which peers are pending — the sweep query gets the
+// authoritative list straight from the database every time. We only
+// need to know the latest disconnect we've observed for this account
+// (so we can decide when it's safe to drop the entry) and the timer
+// that will fire the next sweep.
+type accountEntry struct {
+	lastDisconnectedAt time.Time
+	timer              *time.Timer
 }
 
-// todo: consider to remove peer from ephemeral list when the peer has been deleted via API. If we do not do it
-// in worst case we will get invalid error message in this manager.
-
-// EphemeralManager keep a list of ephemeral peers. After EphemeralLifeTime inactivity the peer will be deleted
-// automatically. Inactivity means the peer disconnected from the Management server.
+// EphemeralManager tracks accounts that may have ephemeral peers in
+// need of cleanup and runs a per-account sweep at the appropriate
+// time. State is in-memory and account-scoped: a sweep deletes any
+// ephemeral peer in the account that has been disconnected for at
+// least lifeTime, then either drops the account from the tracker
+// (when no recent disconnects have arrived) or re-arms the timer.
 type EphemeralManager struct {
 	store        store.Store
 	peersManager peers.Manager
 
-	headPeer  *ephemeralPeer
-	tailPeer  *ephemeralPeer
-	peersLock sync.Mutex
-	timer     *time.Timer
+	accountsLock sync.Mutex
+	accounts     map[string]*accountEntry
+
+	// initialLoadTimer is the one-shot timer used to defer the
+	// post-restart catch-up query; held so Stop() can cancel it.
+	initialLoadTimer *time.Timer
+	// stopped is flipped by Stop() so any timer that fires after
+	// teardown becomes a no-op instead of touching a half-dismantled
+	// store.
+	stopped bool
 
 	lifeTime      time.Duration
 	cleanupWindow time.Duration
 
+	// initialLoadDelay returns the wall-clock delay to wait before
+	// running the post-restart catch-up query. Pluggable so tests can
+	// fire the load immediately.
+	initialLoadDelay func() time.Duration
+
+	// bgCtx is the long-lived context captured at LoadInitialPeers
+	// time. Timer-driven sweeps use it because they fire long after
+	// the original gRPC handler ctx that produced an OnPeerDisconnected
+	// call has been cancelled.
+	bgCtx context.Context
+
 	// metrics is nil-safe; methods on telemetry.EphemeralPeersMetrics
 	// no-op when the receiver is nil so deployments without an app
 	// metrics provider work unchanged.
@@ -58,228 +91,265 @@ type EphemeralManager struct {
 // NewEphemeralManager instantiate new EphemeralManager
 func NewEphemeralManager(store store.Store, peersManager peers.Manager) *EphemeralManager {
 	return &EphemeralManager{
-		store:        store,
-		peersManager: peersManager,
-
-		lifeTime:      ephemeral.EphemeralLifeTime,
-		cleanupWindow: cleanupWindow,
+		store:            store,
+		peersManager:     peersManager,
+		accounts:         make(map[string]*accountEntry),
+		lifeTime:         ephemeral.EphemeralLifeTime,
+		cleanupWindow:    cleanupWindow,
+		initialLoadDelay: defaultInitialLoadDelay,
 	}
 }
 
-// SetMetrics attaches a metrics collector. Safe to call once before
-// LoadInitialPeers; later attachment is fine but earlier loads won't be
-// reflected in the gauge. Pass nil to detach.
+// SetMetrics attaches a metrics collector. Pass nil to detach.
 func (e *EphemeralManager) SetMetrics(m *telemetry.EphemeralPeersMetrics) {
-	e.peersLock.Lock()
+	e.accountsLock.Lock()
 	e.metrics = m
-	e.peersLock.Unlock()
+	e.accountsLock.Unlock()
 }
 
-// LoadInitialPeers load from the database the ephemeral type of peers and schedule a cleanup procedure to the head
-// of the linked list (to the most deprecated peer). At the end of cleanup it schedules the next cleanup to the new
-// head.
+// LoadInitialPeers schedules the post-restart catch-up query for a
+// random moment 8-10 minutes from now. Returns immediately. The
+// catch-up populates the per-account tracker from the database so any
+// peers that disconnected before the restart still get cleaned up.
+//
+// The random delay is critical: without it, every management replica
+// hitting the same Postgres instance after a deploy would issue the
+// catch-up query simultaneously.
 func (e *EphemeralManager) LoadInitialPeers(ctx context.Context) {
-	e.peersLock.Lock()
-	defer e.peersLock.Unlock()
-
-	e.loadEphemeralPeers(ctx)
-	if e.headPeer != nil {
-		e.timer = time.AfterFunc(e.lifeTime, func() {
-			e.cleanup(ctx)
-		})
-	}
-}
-
-// Stop timer
-func (e *EphemeralManager) Stop() {
-	e.peersLock.Lock()
-	defer e.peersLock.Unlock()
-
-	if e.timer != nil {
-		e.timer.Stop()
-	}
-}
-
-// OnPeerConnected remove the peer from the linked list of ephemeral peers. Because it has been called when the peer
-// is active the manager will not delete it while it is active.
-func (e *EphemeralManager) OnPeerConnected(ctx context.Context, peer *nbpeer.Peer) {
-	if !peer.Ephemeral {
+	e.accountsLock.Lock()
+	defer e.accountsLock.Unlock()
+	if e.stopped {
 		return
 	}
 
-	log.WithContext(ctx).Tracef("remove peer from ephemeral list: %s", peer.ID)
+	e.bgCtx = ctx
 
-	e.peersLock.Lock()
-	defer e.peersLock.Unlock()
-
-	if e.removePeer(peer.ID) {
-		e.metrics.DecPending(1)
-	}
-
-	// stop the unnecessary timer
-	if e.headPeer == nil && e.timer != nil {
-		e.timer.Stop()
-		e.timer = nil
-	}
+	delay := e.initialLoadDelay()
+	log.WithContext(ctx).Infof("ephemeral peer initial load scheduled in %s", delay)
+	e.initialLoadTimer = time.AfterFunc(delay, func() {
+		e.loadInitialAccounts(e.bgCtx)
+	})
 }
 
-// OnPeerDisconnected add the peer to the linked list of ephemeral peers. Because of the peer
-// is inactive it will be deleted after the EphemeralLifeTime period.
+// Stop cancels the deferred initial load and any per-account timers.
+func (e *EphemeralManager) Stop() {
+	e.accountsLock.Lock()
+	defer e.accountsLock.Unlock()
+
+	e.stopped = true
+	if e.initialLoadTimer != nil {
+		e.initialLoadTimer.Stop()
+		e.initialLoadTimer = nil
+	}
+	for _, entry := range e.accounts {
+		if entry.timer != nil {
+			entry.timer.Stop()
+		}
+	}
+	e.accounts = make(map[string]*accountEntry)
+}
+
+// OnPeerConnected is a no-op in the account-scoped design. The sweep
+// query filters out connected peers at the database level, so we don't
+// need an explicit "remove from list" signal when a peer reconnects.
+// Kept on the interface to preserve the existing call sites.
+func (e *EphemeralManager) OnPeerConnected(_ context.Context, _ *nbpeer.Peer) {
+}
+
+// OnPeerDisconnected registers a disconnect for the peer's account and
+// arms a sweep if one isn't already scheduled. Non-ephemeral peers are
+// ignored.
 func (e *EphemeralManager) OnPeerDisconnected(ctx context.Context, peer *nbpeer.Peer) {
 	if !peer.Ephemeral {
 		return
 	}
 
-	log.WithContext(ctx).Tracef("add peer to ephemeral list: %s", peer.ID)
-
-	e.peersLock.Lock()
-	defer e.peersLock.Unlock()
-
-	if e.isPeerOnList(peer.ID) {
-		return
-	}
-
-	e.addPeer(peer.AccountID, peer.ID, e.newDeadLine())
-	e.metrics.IncPending()
-	if e.timer == nil {
-		delay := e.headPeer.deadline.Sub(timeNow()) + e.cleanupWindow
-		if delay < 0 {
-			delay = 0
-		}
-		e.timer = time.AfterFunc(delay, func() {
-			e.cleanup(ctx)
-		})
-	}
-}
-
-func (e *EphemeralManager) loadEphemeralPeers(ctx context.Context) {
-	peers, err := e.store.GetAllEphemeralPeers(ctx, store.LockingStrengthNone)
-	if err != nil {
-		log.WithContext(ctx).Debugf("failed to load ephemeral peers: %s", err)
-		return
-	}
-
-	t := e.newDeadLine()
-	for _, p := range peers {
-		e.addPeer(p.AccountID, p.ID, t)
-	}
-	e.metrics.AddPending(int64(len(peers)))
-
-	log.WithContext(ctx).Debugf("loaded ephemeral peer(s): %d", len(peers))
-}
-
-func (e *EphemeralManager) cleanup(ctx context.Context) {
-	log.Tracef("on ephemeral cleanup")
-	deletePeers := make(map[string]*ephemeralPeer)
-
-	e.peersLock.Lock()
 	now := timeNow()
-	for p := e.headPeer; p != nil; p = p.next {
-		if now.Before(p.deadline) {
-			break
-		}
 
-		deletePeers[p.id] = p
-		e.headPeer = p.next
-		if p.next == nil {
-			e.tailPeer = nil
-		}
+	e.accountsLock.Lock()
+	defer e.accountsLock.Unlock()
+	if e.stopped {
+		return
 	}
 
-	if e.headPeer != nil {
-		delay := e.headPeer.deadline.Sub(timeNow()) + e.cleanupWindow
-		if delay < 0 {
-			delay = 0
-		}
-		e.timer = time.AfterFunc(delay, func() {
-			e.cleanup(ctx)
+	entry, existed := e.accounts[peer.AccountID]
+	if !existed {
+		entry = &accountEntry{}
+		e.accounts[peer.AccountID] = entry
+		e.metrics.IncPending()
+	}
+	entry.lastDisconnectedAt = now
+
+	if entry.timer == nil {
+		delay := e.lifeTime + e.cleanupWindow
+		log.WithContext(ctx).Tracef("ephemeral: scheduling sweep for account %s in %s", peer.AccountID, delay)
+		accountID := peer.AccountID
+		entry.timer = time.AfterFunc(delay, func() {
+			e.sweep(e.bgCtxOrFallback(ctx), accountID)
 		})
-	} else {
-		e.timer = nil
+	}
+}
+
+// bgCtxOrFallback returns the long-lived background context captured at
+// LoadInitialPeers time, falling back to the supplied ctx when the
+// manager hasn't been started through LoadInitialPeers (e.g. in tests
+// that drive the manager directly). Must be called with the lock held
+// or before the timer is armed.
+func (e *EphemeralManager) bgCtxOrFallback(ctx context.Context) context.Context {
+	if e.bgCtx != nil {
+		return e.bgCtx
+	}
+	return ctx
+}
+
+// loadInitialAccounts runs the post-restart catch-up query and seeds
+// the tracker with one entry per account that has at least one
+// disconnected ephemeral peer.
+func (e *EphemeralManager) loadInitialAccounts(ctx context.Context) {
+	accounts, err := e.store.GetEphemeralAccountsLastDisconnect(ctx)
+	if err != nil {
+		log.WithContext(ctx).Errorf("failed to load ephemeral accounts on startup: %v", err)
+		return
 	}
 
-	e.peersLock.Unlock()
+	now := timeNow()
+	added := 0
 
-	// Drop the gauge by the number of entries we just took off the list,
-	// regardless of whether the subsequent DeletePeers call succeeds. The
-	// list invariant is what the gauge tracks; failed delete batches are
-	// counted separately via CountCleanupError so we can still see them.
-	if len(deletePeers) > 0 {
-		e.metrics.CountCleanupRun()
-		e.metrics.DecPending(int64(len(deletePeers)))
+	e.accountsLock.Lock()
+	defer e.accountsLock.Unlock()
+	if e.stopped {
+		return
 	}
 
-	peerIDsPerAccount := make(map[string][]string)
-	for id, p := range deletePeers {
-		peerIDsPerAccount[p.accountID] = append(peerIDsPerAccount[p.accountID], id)
-	}
-
-	for accountID, peerIDs := range peerIDsPerAccount {
-		log.WithContext(ctx).Tracef("cleanup: deleting %d ephemeral peers for account %s", len(peerIDs), accountID)
-		err := e.peersManager.DeletePeers(ctx, accountID, peerIDs, activity.SystemInitiator, true)
-		if err != nil {
-			log.WithContext(ctx).Errorf("failed to delete ephemeral peers: %s", err)
-			e.metrics.CountCleanupError()
+	for accountID, lastDisc := range accounts {
+		// If we already learned about this account via an
+		// OnPeerDisconnected that arrived during the random delay
+		// window, prefer the live timestamp.
+		if _, alreadyTracked := e.accounts[accountID]; alreadyTracked {
 			continue
 		}
-		e.metrics.CountPeersCleaned(int64(len(peerIDs)))
-	}
-}
 
-func (e *EphemeralManager) addPeer(accountID string, peerID string, deadline time.Time) {
-	ep := &ephemeralPeer{
-		id:        peerID,
-		accountID: accountID,
-		deadline:  deadline,
-	}
+		entry := &accountEntry{lastDisconnectedAt: lastDisc}
+		horizon := lastDisc.Add(e.lifeTime)
 
-	if e.headPeer == nil {
-		e.headPeer = ep
-	}
-	if e.tailPeer != nil {
-		e.tailPeer.next = ep
-	}
-	e.tailPeer = ep
-}
-
-// removePeer drops the entry from the linked list. Returns true if a
-// matching entry was found and removed so callers can keep the pending
-// metric gauge in sync.
-func (e *EphemeralManager) removePeer(id string) bool {
-	if e.headPeer == nil {
-		return false
-	}
-
-	if e.headPeer.id == id {
-		e.headPeer = e.headPeer.next
-		if e.tailPeer.id == id {
-			e.tailPeer = nil
+		var delay time.Duration
+		if horizon.After(now) {
+			delay = horizon.Sub(now) + e.cleanupWindow
+		} else {
+			// Already past the staleness window — sweep right away
+			// (one cleanupWindow later, to keep startup load smooth
+			// when many accounts qualify at once).
+			delay = e.cleanupWindow
 		}
-		return true
+		idForClosure := accountID
+		entry.timer = time.AfterFunc(delay, func() {
+			e.sweep(ctx, idForClosure)
+		})
+		e.accounts[accountID] = entry
+		added++
 	}
 
-	for p := e.headPeer; p.next != nil; p = p.next {
-		if p.next.id == id {
-			// if we remove the last element from the chain then set the last-1 as tail
-			if e.tailPeer.id == id {
-				e.tailPeer = p
-			}
-			p.next = p.next.next
-			return true
+	e.metrics.AddPending(int64(added))
+	log.WithContext(ctx).Debugf("ephemeral: loaded %d account(s) for cleanup tracking", added)
+}
+
+// sweep runs the cleanup pass for a single account. It queries the
+// database for disconnected ephemeral peers that have crossed the
+// staleness window, deletes them via peers.Manager, and then decides
+// whether to drop the account from the tracker or re-arm the timer.
+func (e *EphemeralManager) sweep(ctx context.Context, accountID string) {
+	now := timeNow()
+
+	e.accountsLock.Lock()
+	entry, ok := e.accounts[accountID]
+	if !ok || e.stopped {
+		e.accountsLock.Unlock()
+		return
+	}
+	lastDisc := entry.lastDisconnectedAt
+	entry.timer = nil
+	e.accountsLock.Unlock()
+
+	threshold := now.Add(-e.lifeTime)
+	stalePeerIDs, err := e.store.GetStaleEphemeralPeerIDsForAccount(ctx, accountID, threshold)
+	if err != nil {
+		log.WithContext(ctx).Errorf("ephemeral: failed to query stale peers for account %s: %v", accountID, err)
+		e.metrics.CountCleanupError()
+		e.rearm(ctx, accountID, e.cleanupWindow)
+		return
+	}
+
+	if len(stalePeerIDs) > 0 {
+		log.WithContext(ctx).Tracef("ephemeral: deleting %d peer(s) for account %s", len(stalePeerIDs), accountID)
+		if err := e.peersManager.DeletePeers(ctx, accountID, stalePeerIDs, activity.SystemInitiator, true); err != nil {
+			log.WithContext(ctx).Errorf("ephemeral: failed to delete peers for account %s: %v", accountID, err)
+			e.metrics.CountCleanupError()
+			e.rearm(ctx, accountID, e.cleanupWindow)
+			return
 		}
+		e.metrics.CountCleanupRun()
+		e.metrics.CountPeersCleaned(int64(len(stalePeerIDs)))
 	}
-	return false
+
+	e.accountsLock.Lock()
+	defer e.accountsLock.Unlock()
+	if e.stopped {
+		return
+	}
+	entry, ok = e.accounts[accountID]
+	if !ok {
+		return
+	}
+
+	// Drop rule: if every disconnect we've observed has now crossed
+	// the staleness window, the sweep we just ran saw everything that
+	// could possibly need cleaning. Dropping is safe — a future
+	// disconnect will recreate the entry. The check uses the latest
+	// lastDisc, which may have advanced (concurrently with the sweep
+	// itself) due to a new OnPeerDisconnected, in which case we
+	// correctly re-arm.
+	horizon := entry.lastDisconnectedAt.Add(e.lifeTime)
+	if !horizon.After(now) {
+		delete(e.accounts, accountID)
+		e.metrics.DecPending(1)
+		log.WithContext(ctx).Tracef("ephemeral: dropping account %s (lastDisc=%s, horizon=%s, now=%s)",
+			accountID, lastDisc, horizon, now)
+		return
+	}
+
+	delay := horizon.Sub(now) + e.cleanupWindow
+	idForClosure := accountID
+	entry.timer = time.AfterFunc(delay, func() {
+		e.sweep(ctx, idForClosure)
+	})
 }
 
-func (e *EphemeralManager) isPeerOnList(id string) bool {
-	for p := e.headPeer; p != nil; p = p.next {
-		if p.id == id {
-			return true
-		}
+// rearm reschedules a sweep `delay` from now. Used after a recoverable
+// error in the sweep path so the account doesn't get stuck.
+func (e *EphemeralManager) rearm(ctx context.Context, accountID string, delay time.Duration) {
+	e.accountsLock.Lock()
+	defer e.accountsLock.Unlock()
+	if e.stopped {
+		return
 	}
-	return false
+	entry, ok := e.accounts[accountID]
+	if !ok {
+		return
+	}
+	idForClosure := accountID
+	entry.timer = time.AfterFunc(delay, func() {
+		e.sweep(ctx, idForClosure)
+	})
 }
 
-func (e *EphemeralManager) newDeadLine() time.Time {
-	return timeNow().Add(e.lifeTime)
+// defaultInitialLoadDelay returns a random duration in
+// [initialLoadMinDelay, initialLoadMaxDelay). Process-wide
+// math/rand is acceptable here — the delay is purely a smoothing
+// jitter, not a security primitive.
+func defaultInitialLoadDelay() time.Duration {
+	span := int64(initialLoadMaxDelay - initialLoadMinDelay)
+	if span <= 0 {
+		return initialLoadMinDelay
+	}
+	return initialLoadMinDelay + time.Duration(rand.Int63n(span))
 }
diff --git a/management/internals/modules/peers/ephemeral/manager/ephemeral_test.go b/management/internals/modules/peers/ephemeral/manager/ephemeral_test.go
index 314e84501..65a495e7d 100644
--- a/management/internals/modules/peers/ephemeral/manager/ephemeral_test.go
+++ b/management/internals/modules/peers/ephemeral/manager/ephemeral_test.go
@@ -2,299 +2,544 @@ package manager
 
 import (
 	"context"
+	"errors"
 	"fmt"
 	"sync"
+	"sync/atomic"
 	"testing"
 	"time"
 
 	"github.com/golang/mock/gomock"
 	log "github.com/sirupsen/logrus"
 	"github.com/stretchr/testify/assert"
+	"github.com/stretchr/testify/require"
 
 	nbdns "github.com/netbirdio/netbird/dns"
 	"github.com/netbirdio/netbird/management/internals/modules/peers"
 	"github.com/netbirdio/netbird/management/internals/modules/peers/ephemeral"
-	nbAccount "github.com/netbirdio/netbird/management/server/account"
 	nbpeer "github.com/netbirdio/netbird/management/server/peer"
 	"github.com/netbirdio/netbird/management/server/store"
 	"github.com/netbirdio/netbird/management/server/types"
 	"github.com/netbirdio/netbird/route"
 )
 
+// MockStore is a thin in-memory stand-in that implements only the two
+// methods the EphemeralManager uses. It honors the account / ephemeral
+// / connected / lastSeen attributes of each peer so the cleanup logic
+// can be exercised end-to-end without bringing up sqlite or Postgres.
 type MockStore struct {
 	store.Store
+	mu      sync.Mutex
 	account *types.Account
 }
 
-func (s *MockStore) GetAllEphemeralPeers(_ context.Context, _ store.LockingStrength) ([]*nbpeer.Peer, error) {
-	var peers []*nbpeer.Peer
-	for _, v := range s.account.Peers {
-		if v.Ephemeral {
-			peers = append(peers, v)
+func (s *MockStore) GetStaleEphemeralPeerIDsForAccount(_ context.Context, accountID string, olderThan time.Time) ([]string, error) {
+	s.mu.Lock()
+	defer s.mu.Unlock()
+	if s.account == nil || s.account.Id != accountID {
+		return nil, nil
+	}
+	var ids []string
+	for _, p := range s.account.Peers {
+		if !p.Ephemeral {
+			continue
+		}
+		if p.Status == nil || p.Status.Connected {
+			continue
+		}
+		if p.Status.LastSeen.Before(olderThan) {
+			ids = append(ids, p.ID)
 		}
 	}
-	return peers, nil
+	return ids, nil
 }
 
-type MockAccountManager struct {
-	mu sync.Mutex
-	nbAccount.Manager
-	store             *MockStore
-	deletePeerCalls   int
-	bufferUpdateCalls map[string]int
-	wg                *sync.WaitGroup
-}
-
-func (a *MockAccountManager) DeletePeer(_ context.Context, accountID, peerID, userID string) error {
-	a.mu.Lock()
-	defer a.mu.Unlock()
-	a.deletePeerCalls++
-	delete(a.store.account.Peers, peerID)
-	if a.wg != nil {
-		a.wg.Done()
+func (s *MockStore) GetEphemeralAccountsLastDisconnect(_ context.Context) (map[string]time.Time, error) {
+	s.mu.Lock()
+	defer s.mu.Unlock()
+	out := map[string]time.Time{}
+	if s.account == nil {
+		return out, nil
 	}
-	return nil
-}
-
-func (a *MockAccountManager) GetDeletePeerCalls() int {
-	a.mu.Lock()
-	defer a.mu.Unlock()
-	return a.deletePeerCalls
-}
-
-func (a *MockAccountManager) BufferUpdateAccountPeers(ctx context.Context, accountID string, reason types.UpdateReason) {
-	a.mu.Lock()
-	defer a.mu.Unlock()
-	if a.bufferUpdateCalls == nil {
-		a.bufferUpdateCalls = make(map[string]int)
+	var latest time.Time
+	hasAny := false
+	for _, p := range s.account.Peers {
+		if !p.Ephemeral || p.Status == nil || p.Status.Connected {
+			continue
+		}
+		if !hasAny || p.Status.LastSeen.After(latest) {
+			latest = p.Status.LastSeen
+			hasAny = true
+		}
 	}
-	a.bufferUpdateCalls[accountID]++
-}
-
-func (a *MockAccountManager) GetBufferUpdateCalls(accountID string) int {
-	a.mu.Lock()
-	defer a.mu.Unlock()
-	if a.bufferUpdateCalls == nil {
-		return 0
+	if hasAny {
+		out[s.account.Id] = latest
 	}
-	return a.bufferUpdateCalls[accountID]
+	return out, nil
 }
 
-func (a *MockAccountManager) GetStore() store.Store {
-	return a.store
-}
-
-func TestNewManager(t *testing.T) {
-	t.Cleanup(func() {
-		timeNow = time.Now
-	})
-	startTime := time.Now()
+// withFakeClock pins timeNow to a settable value for the duration of t.
+// Returns a getter and a setter so subtests can advance virtual time.
+func withFakeClock(t *testing.T, start time.Time) (get func() time.Time, set func(time.Time)) {
+	t.Helper()
+	var mu sync.Mutex
+	now := start
 	timeNow = func() time.Time {
-		return startTime
+		mu.Lock()
+		defer mu.Unlock()
+		return now
 	}
+	t.Cleanup(func() { timeNow = time.Now })
 
-	store := &MockStore{}
-	ctrl := gomock.NewController(t)
-	peersManager := peers.NewMockManager(ctrl)
-
-	numberOfPeers := 5
-	numberOfEphemeralPeers := 3
-	seedPeers(store, numberOfPeers, numberOfEphemeralPeers)
-
-	// Expect DeletePeers to be called for ephemeral peers
-	peersManager.EXPECT().
-		DeletePeers(gomock.Any(), gomock.Any(), gomock.Any(), gomock.Any(), true).
-		DoAndReturn(func(ctx context.Context, accountID string, peerIDs []string, userID string, checkConnected bool) error {
-			for _, peerID := range peerIDs {
-				delete(store.account.Peers, peerID)
-			}
-			return nil
-		}).
-		AnyTimes()
-
-	mgr := NewEphemeralManager(store, peersManager)
-	mgr.loadEphemeralPeers(context.Background())
-	startTime = startTime.Add(ephemeral.EphemeralLifeTime + 1)
-	mgr.cleanup(context.Background())
-
-	if len(store.account.Peers) != numberOfPeers {
-		t.Errorf("failed to cleanup ephemeral peers, expected: %d, result: %d", numberOfPeers, len(store.account.Peers))
-	}
+	return func() time.Time {
+			mu.Lock()
+			defer mu.Unlock()
+			return now
+		}, func(v time.Time) {
+			mu.Lock()
+			defer mu.Unlock()
+			now = v
+		}
 }
 
-func TestNewManagerPeerConnected(t *testing.T) {
-	t.Cleanup(func() {
-		timeNow = time.Now
-	})
-	startTime := time.Now()
-	timeNow = func() time.Time {
-		return startTime
-	}
-
-	store := &MockStore{}
-	ctrl := gomock.NewController(t)
-	peersManager := peers.NewMockManager(ctrl)
-
-	numberOfPeers := 5
-	numberOfEphemeralPeers := 3
-	seedPeers(store, numberOfPeers, numberOfEphemeralPeers)
-
-	// Expect DeletePeers to be called for ephemeral peers (except the connected one)
-	peersManager.EXPECT().
-		DeletePeers(gomock.Any(), gomock.Any(), gomock.Any(), gomock.Any(), true).
-		DoAndReturn(func(ctx context.Context, accountID string, peerIDs []string, userID string, checkConnected bool) error {
-			for _, peerID := range peerIDs {
-				delete(store.account.Peers, peerID)
-			}
-			return nil
-		}).
-		AnyTimes()
-
-	mgr := NewEphemeralManager(store, peersManager)
-	mgr.loadEphemeralPeers(context.Background())
-	mgr.OnPeerConnected(context.Background(), store.account.Peers["ephemeral_peer_0"])
-
-	startTime = startTime.Add(ephemeral.EphemeralLifeTime + 1)
-	mgr.cleanup(context.Background())
-
-	expected := numberOfPeers + 1
-	if len(store.account.Peers) != expected {
-		t.Errorf("failed to cleanup ephemeral peers, expected: %d, result: %d", expected, len(store.account.Peers))
-	}
+// newManagerForTest builds a manager with short timers and no random
+// initial-load delay so tests run instantly.
+func newManagerForTest(t *testing.T, st store.Store, peersMgr peers.Manager) *EphemeralManager {
+	t.Helper()
+	mgr := NewEphemeralManager(st, peersMgr)
+	mgr.lifeTime = 100 * time.Millisecond
+	mgr.cleanupWindow = 10 * time.Millisecond
+	mgr.initialLoadDelay = func() time.Duration { return 0 }
+	t.Cleanup(mgr.Stop)
+	return mgr
 }
 
-func TestNewManagerPeerDisconnected(t *testing.T) {
-	t.Cleanup(func() {
-		timeNow = time.Now
-	})
-	startTime := time.Now()
-	timeNow = func() time.Time {
-		return startTime
-	}
+// TestOnPeerDisconnected_RegistersAndSweeps drives the OnPeerDisconnected
+// path with a fake clock: a single ephemeral peer disconnects, we
+// advance past the staleness window, and the sweep deletes it.
+func TestOnPeerDisconnected_RegistersAndSweeps(t *testing.T) {
+	mockStore := &MockStore{account: newAccountWithId(context.Background(), "acc-1", "", "", false)}
 
-	store := &MockStore{}
-	ctrl := gomock.NewController(t)
-	peersManager := peers.NewMockManager(ctrl)
-
-	numberOfPeers := 5
-	numberOfEphemeralPeers := 3
-	seedPeers(store, numberOfPeers, numberOfEphemeralPeers)
-
-	// Expect DeletePeers to be called for the one disconnected peer
-	peersManager.EXPECT().
-		DeletePeers(gomock.Any(), gomock.Any(), gomock.Any(), gomock.Any(), true).
-		DoAndReturn(func(ctx context.Context, accountID string, peerIDs []string, userID string, checkConnected bool) error {
-			for _, peerID := range peerIDs {
-				delete(store.account.Peers, peerID)
-			}
-			return nil
-		}).
-		AnyTimes()
-
-	mgr := NewEphemeralManager(store, peersManager)
-	mgr.loadEphemeralPeers(context.Background())
-	for _, v := range store.account.Peers {
-		mgr.OnPeerConnected(context.Background(), v)
-
-	}
-	mgr.OnPeerDisconnected(context.Background(), store.account.Peers["ephemeral_peer_0"])
-
-	startTime = startTime.Add(ephemeral.EphemeralLifeTime + 1)
-	mgr.cleanup(context.Background())
-
-	expected := numberOfPeers + numberOfEphemeralPeers - 1
-	if len(store.account.Peers) != expected {
-		t.Errorf("failed to cleanup ephemeral peers, expected: %d, result: %d", expected, len(store.account.Peers))
-	}
-}
-
-func TestCleanupSchedulingBehaviorIsBatched(t *testing.T) {
-	const (
-		ephemeralPeers    = 10
-		testLifeTime      = 1 * time.Second
-		testCleanupWindow = 100 * time.Millisecond
-	)
-
-	t.Cleanup(func() {
-		timeNow = time.Now
-	})
-	startTime := time.Now()
-	timeNow = func() time.Time {
-		return startTime
-	}
-
-	mockStore := &MockStore{}
-	account := newAccountWithId(context.Background(), "account", "", "", false)
-	mockStore.account = account
-
-	wg := &sync.WaitGroup{}
-	wg.Add(ephemeralPeers)
-	mockAM := &MockAccountManager{
-		store: mockStore,
-		wg:    wg,
-	}
+	getNow, setNow := withFakeClock(t, time.Now())
 
 	ctrl := gomock.NewController(t)
-	peersManager := peers.NewMockManager(ctrl)
+	peersMgr := peers.NewMockManager(ctrl)
 
-	// Set up expectation that DeletePeers will be called once with all peer IDs
-	peersManager.EXPECT().
-		DeletePeers(gomock.Any(), account.Id, gomock.Any(), gomock.Any(), true).
-		DoAndReturn(func(ctx context.Context, accountID string, peerIDs []string, userID string, checkConnected bool) error {
-			// Simulate the actual deletion behavior
-			for _, peerID := range peerIDs {
-				err := mockAM.DeletePeer(ctx, accountID, peerID, userID)
-				if err != nil {
-					return err
-				}
+	var deletedMu sync.Mutex
+	var deleted []string
+	var deleteCalls atomic.Int32
+	peersMgr.EXPECT().
+		DeletePeers(gomock.Any(), "acc-1", gomock.Any(), gomock.Any(), true).
+		DoAndReturn(func(_ context.Context, accountID string, peerIDs []string, _ string, _ bool) error {
+			deleteCalls.Add(1)
+			mockStore.mu.Lock()
+			for _, id := range peerIDs {
+				delete(mockStore.account.Peers, id)
 			}
-			mockAM.BufferUpdateAccountPeers(ctx, accountID, types.UpdateReason{})
+			mockStore.mu.Unlock()
+			deletedMu.Lock()
+			deleted = append(deleted, peerIDs...)
+			deletedMu.Unlock()
 			return nil
-		}).
-		Times(1)
+		}).AnyTimes()
 
-	mgr := NewEphemeralManager(mockStore, peersManager)
-	mgr.lifeTime = testLifeTime
-	mgr.cleanupWindow = testCleanupWindow
+	mgr := newManagerForTest(t, mockStore, peersMgr)
 
-	// Add peers and disconnect them at slightly different times (within cleanup window)
-	for i := range ephemeralPeers {
-		p := &nbpeer.Peer{ID: fmt.Sprintf("peer-%d", i), AccountID: account.Id, Ephemeral: true}
-		mockStore.account.Peers[p.ID] = p
+	// One ephemeral peer that disconnected "now".
+	now := getNow()
+	p := &nbpeer.Peer{
+		ID:        "p1",
+		AccountID: "acc-1",
+		Ephemeral: true,
+		Status:    &nbpeer.PeerStatus{Connected: false, LastSeen: now},
+	}
+	mockStore.account.Peers[p.ID] = p
+	mgr.OnPeerDisconnected(context.Background(), p)
+
+	// Advance past lifeTime + cleanupWindow so the timer-driven sweep fires.
+	setNow(now.Add(mgr.lifeTime + 5*mgr.cleanupWindow))
+	require.Eventually(t, func() bool { return deleteCalls.Load() >= 1 }, 2*time.Second, 5*time.Millisecond,
+		"sweep should fire and delete the stale peer")
+
+	deletedMu.Lock()
+	deletedCopy := append([]string(nil), deleted...)
+	deletedMu.Unlock()
+	require.Equal(t, []string{"p1"}, deletedCopy, "only the one ephemeral peer should be deleted")
+}
+
+// TestOnPeerDisconnected_NonEphemeralIgnored: a non-ephemeral disconnect
+// must not register the account or arm any timer.
+func TestOnPeerDisconnected_NonEphemeralIgnored(t *testing.T) {
+	mockStore := &MockStore{account: newAccountWithId(context.Background(), "acc-1", "", "", false)}
+	withFakeClock(t, time.Now())
+
+	ctrl := gomock.NewController(t)
+	peersMgr := peers.NewMockManager(ctrl)
+	// No DeletePeers expectation — must not be called.
+
+	mgr := newManagerForTest(t, mockStore, peersMgr)
+	mgr.OnPeerDisconnected(context.Background(), &nbpeer.Peer{
+		ID:        "p1",
+		AccountID: "acc-1",
+		Ephemeral: false,
+		Status:    &nbpeer.PeerStatus{Connected: false, LastSeen: timeNow()},
+	})
+
+	mgr.accountsLock.Lock()
+	require.Empty(t, mgr.accounts, "non-ephemeral disconnect must not register an account")
+	mgr.accountsLock.Unlock()
+}
+
+// TestSweep_DropsAccountWhenIdle: after a sweep cleans the stale peers,
+// if no more disconnects have arrived the account must be dropped from
+// the in-memory tracker.
+func TestSweep_DropsAccountWhenIdle(t *testing.T) {
+	mockStore := &MockStore{account: newAccountWithId(context.Background(), "acc-1", "", "", false)}
+	getNow, setNow := withFakeClock(t, time.Now())
+
+	ctrl := gomock.NewController(t)
+	peersMgr := peers.NewMockManager(ctrl)
+	peersMgr.EXPECT().
+		DeletePeers(gomock.Any(), "acc-1", gomock.Any(), gomock.Any(), true).
+		DoAndReturn(func(_ context.Context, _ string, peerIDs []string, _ string, _ bool) error {
+			mockStore.mu.Lock()
+			for _, id := range peerIDs {
+				delete(mockStore.account.Peers, id)
+			}
+			mockStore.mu.Unlock()
+			return nil
+		}).AnyTimes()
+
+	mgr := newManagerForTest(t, mockStore, peersMgr)
+
+	now := getNow()
+	p := &nbpeer.Peer{ID: "p1", AccountID: "acc-1", Ephemeral: true,
+		Status: &nbpeer.PeerStatus{Connected: false, LastSeen: now}}
+	mockStore.account.Peers[p.ID] = p
+	mgr.OnPeerDisconnected(context.Background(), p)
+
+	setNow(now.Add(mgr.lifeTime + 5*mgr.cleanupWindow))
+
+	require.Eventually(t, func() bool {
+		mgr.accountsLock.Lock()
+		defer mgr.accountsLock.Unlock()
+		return len(mgr.accounts) == 0
+	}, 2*time.Second, 5*time.Millisecond, "account should be dropped after sweep with no new disconnects")
+}
+
+// TestSweep_ReArmsWhenNewDisconnectArrived: simulate the race where a
+// fresh disconnect arrives just before the sweep fires. The sweep must
+// observe the updated lastDisc and re-arm rather than drop.
+func TestSweep_ReArmsWhenNewDisconnectArrived(t *testing.T) {
+	mockStore := &MockStore{account: newAccountWithId(context.Background(), "acc-1", "", "", false)}
+	getNow, setNow := withFakeClock(t, time.Now())
+
+	ctrl := gomock.NewController(t)
+	peersMgr := peers.NewMockManager(ctrl)
+	peersMgr.EXPECT().
+		DeletePeers(gomock.Any(), "acc-1", gomock.Any(), gomock.Any(), true).
+		DoAndReturn(func(_ context.Context, _ string, peerIDs []string, _ string, _ bool) error {
+			mockStore.mu.Lock()
+			for _, id := range peerIDs {
+				delete(mockStore.account.Peers, id)
+			}
+			mockStore.mu.Unlock()
+			return nil
+		}).AnyTimes()
+
+	mgr := newManagerForTest(t, mockStore, peersMgr)
+
+	now := getNow()
+	p1 := &nbpeer.Peer{ID: "p1", AccountID: "acc-1", Ephemeral: true,
+		Status: &nbpeer.PeerStatus{Connected: false, LastSeen: now}}
+	mockStore.account.Peers[p1.ID] = p1
+	mgr.OnPeerDisconnected(context.Background(), p1)
+
+	// Advance most of the way toward the first sweep, then introduce
+	// a fresh disconnect that resets lastDisc.
+	setNow(now.Add(mgr.lifeTime - 10*time.Millisecond))
+	p2 := &nbpeer.Peer{ID: "p2", AccountID: "acc-1", Ephemeral: true,
+		Status: &nbpeer.PeerStatus{Connected: false, LastSeen: getNow()}}
+	mockStore.account.Peers[p2.ID] = p2
+	mgr.OnPeerDisconnected(context.Background(), p2)
+
+	// Push past p1's staleness so the first sweep runs and cleans p1
+	// but observes p2 already on the account entry. It must re-arm.
+	setNow(now.Add(mgr.lifeTime + 5*mgr.cleanupWindow))
+
+	require.Eventually(t, func() bool {
+		mockStore.mu.Lock()
+		defer mockStore.mu.Unlock()
+		_, gone := mockStore.account.Peers["p1"]
+		return !gone
+	}, 2*time.Second, 5*time.Millisecond, "p1 should be cleaned at the first sweep")
+
+	// The account should still be tracked because p2 is younger than lifeTime
+	// from the sweep's vantage point at this moment.
+	mgr.accountsLock.Lock()
+	_, stillTracked := mgr.accounts["acc-1"]
+	mgr.accountsLock.Unlock()
+	require.True(t, stillTracked, "account should remain tracked because p2's disconnect kept it active")
+
+	// Push past p2's staleness; second sweep cleans p2 and drops the account.
+	setNow(getNow().Add(mgr.lifeTime + 5*mgr.cleanupWindow))
+	require.Eventually(t, func() bool {
+		mgr.accountsLock.Lock()
+		defer mgr.accountsLock.Unlock()
+		return len(mgr.accounts) == 0
+	}, 2*time.Second, 5*time.Millisecond, "account should drop after the final sweep")
+}
+
+// TestSweep_BatchesPeersPerAccount: many ephemeral peers disconnect on
+// the same account; a single sweep must delete them all in one
+// DeletePeers call.
+func TestSweep_BatchesPeersPerAccount(t *testing.T) {
+	const ephemeralPeers = 8
+
+	mockStore := &MockStore{account: newAccountWithId(context.Background(), "acc-1", "", "", false)}
+	getNow, setNow := withFakeClock(t, time.Now())
+
+	ctrl := gomock.NewController(t)
+	peersMgr := peers.NewMockManager(ctrl)
+
+	deleteBatches := make(chan []string, 4)
+	peersMgr.EXPECT().
+		DeletePeers(gomock.Any(), "acc-1", gomock.Any(), gomock.Any(), true).
+		DoAndReturn(func(_ context.Context, _ string, peerIDs []string, _ string, _ bool) error {
+			cp := append([]string(nil), peerIDs...)
+			mockStore.mu.Lock()
+			for _, id := range peerIDs {
+				delete(mockStore.account.Peers, id)
+			}
+			mockStore.mu.Unlock()
+			deleteBatches <- cp
+			return nil
+		}).Times(1)
+
+	mgr := newManagerForTest(t, mockStore, peersMgr)
+
+	now := getNow()
+	for i := 0; i < ephemeralPeers; i++ {
+		id := fmt.Sprintf("p-%d", i)
+		// Stagger by a fraction of cleanupWindow so they all fall on
+		// the same sweep tick.
+		when := now.Add(time.Duration(i) * time.Millisecond)
+		p := &nbpeer.Peer{ID: id, AccountID: "acc-1", Ephemeral: true,
+			Status: &nbpeer.PeerStatus{Connected: false, LastSeen: when}}
+		mockStore.account.Peers[id] = p
 		mgr.OnPeerDisconnected(context.Background(), p)
-		startTime = startTime.Add(testCleanupWindow / (ephemeralPeers * 2))
 	}
 
-	// Advance time past the lifetime to trigger cleanup
-	startTime = startTime.Add(testLifeTime + testCleanupWindow)
+	setNow(now.Add(mgr.lifeTime + 5*mgr.cleanupWindow))
 
-	// Wait for all deletions to complete
-	wg.Wait()
-
-	assert.Len(t, mockStore.account.Peers, 0, "all ephemeral peers should be cleaned up after the lifetime")
-	assert.Equal(t, 1, mockAM.GetBufferUpdateCalls(account.Id), "buffer update should be called once")
-	assert.Equal(t, ephemeralPeers, mockAM.GetDeletePeerCalls(), "should have deleted all peers")
+	select {
+	case batch := <-deleteBatches:
+		require.Len(t, batch, ephemeralPeers, "all peers should be deleted in a single batch")
+	case <-time.After(2 * time.Second):
+		t.Fatal("expected one batched DeletePeers call")
+	}
 }
 
-func seedPeers(store *MockStore, numberOfPeers int, numberOfEphemeralPeers int) {
-	store.account = newAccountWithId(context.Background(), "my account", "", "", false)
+// TestLoadInitialAccounts_SeedsFromStore exercises the post-restart
+// catch-up path: pre-populate the store, point the manager at it, and
+// confirm both already-stale and not-yet-stale peers get cleaned at
+// their proper times.
+func TestLoadInitialAccounts_SeedsFromStore(t *testing.T) {
+	mockStore := &MockStore{account: newAccountWithId(context.Background(), "acc-1", "", "", false)}
+	getNow, setNow := withFakeClock(t, time.Now())
 
-	for i := 0; i < numberOfPeers; i++ {
-		peerId := fmt.Sprintf("peer_%d", i)
-		p := &nbpeer.Peer{
-			ID:        peerId,
-			Ephemeral: false,
-		}
-		store.account.Peers[p.ID] = p
+	now := getNow()
+	// p-stale: already past the staleness window when load runs.
+	mockStore.account.Peers["p-stale"] = &nbpeer.Peer{
+		ID: "p-stale", AccountID: "acc-1", Ephemeral: true,
+		Status: &nbpeer.PeerStatus{Connected: false, LastSeen: now.Add(-time.Hour)},
+	}
+	// p-fresh: disconnected but not yet stale.
+	mockStore.account.Peers["p-fresh"] = &nbpeer.Peer{
+		ID: "p-fresh", AccountID: "acc-1", Ephemeral: true,
+		Status: &nbpeer.PeerStatus{Connected: false, LastSeen: now},
 	}
 
-	for i := 0; i < numberOfEphemeralPeers; i++ {
-		peerId := fmt.Sprintf("ephemeral_peer_%d", i)
-		p := &nbpeer.Peer{
-			ID:        peerId,
-			Ephemeral: true,
-		}
-		store.account.Peers[p.ID] = p
+	ctrl := gomock.NewController(t)
+	peersMgr := peers.NewMockManager(ctrl)
+	peersMgr.EXPECT().
+		DeletePeers(gomock.Any(), "acc-1", gomock.Any(), gomock.Any(), true).
+		DoAndReturn(func(_ context.Context, _ string, peerIDs []string, _ string, _ bool) error {
+			mockStore.mu.Lock()
+			for _, id := range peerIDs {
+				delete(mockStore.account.Peers, id)
+			}
+			mockStore.mu.Unlock()
+			return nil
+		}).AnyTimes()
+
+	mgr := newManagerForTest(t, mockStore, peersMgr)
+	// Drive loadInitialAccounts directly with the fake-clock-aware now.
+	mgr.loadInitialAccounts(context.Background())
+
+	// First sweep should fire shortly (cleanupWindow) for the stale peer.
+	setNow(now.Add(5 * mgr.cleanupWindow))
+	require.Eventually(t, func() bool {
+		mockStore.mu.Lock()
+		defer mockStore.mu.Unlock()
+		_, gone := mockStore.account.Peers["p-stale"]
+		return !gone
+	}, 2*time.Second, 5*time.Millisecond, "p-stale should be deleted on the first sweep")
+
+	// p-fresh is not yet stale; advance past its window.
+	setNow(now.Add(mgr.lifeTime + 5*mgr.cleanupWindow))
+	require.Eventually(t, func() bool {
+		mockStore.mu.Lock()
+		defer mockStore.mu.Unlock()
+		_, gone := mockStore.account.Peers["p-fresh"]
+		return !gone
+	}, 2*time.Second, 5*time.Millisecond, "p-fresh should be deleted once it crosses the staleness window")
+}
+
+// TestStop_CancelsPendingWork verifies that Stop() cancels both the
+// deferred initial load and per-account sweep timers and that
+// subsequent OnPeerDisconnected calls are ignored.
+func TestStop_CancelsPendingWork(t *testing.T) {
+	mockStore := &MockStore{account: newAccountWithId(context.Background(), "acc-1", "", "", false)}
+	withFakeClock(t, time.Now())
+
+	ctrl := gomock.NewController(t)
+	peersMgr := peers.NewMockManager(ctrl)
+	// DeletePeers must NOT be called after Stop.
+
+	mgr := NewEphemeralManager(mockStore, peersMgr)
+	mgr.lifeTime = 100 * time.Millisecond
+	mgr.cleanupWindow = 10 * time.Millisecond
+	// Use a long delay so the initial-load timer is still pending.
+	mgr.initialLoadDelay = func() time.Duration { return time.Hour }
+
+	mgr.LoadInitialPeers(context.Background())
+	mgr.OnPeerDisconnected(context.Background(), &nbpeer.Peer{
+		ID: "p1", AccountID: "acc-1", Ephemeral: true,
+		Status: &nbpeer.PeerStatus{Connected: false, LastSeen: timeNow()},
+	})
+
+	mgr.accountsLock.Lock()
+	require.NotNil(t, mgr.initialLoadTimer, "initial-load timer should be armed")
+	require.Len(t, mgr.accounts, 1, "account should be tracked after disconnect")
+	mgr.accountsLock.Unlock()
+
+	mgr.Stop()
+
+	mgr.accountsLock.Lock()
+	require.Empty(t, mgr.accounts, "Stop should clear tracked accounts")
+	require.True(t, mgr.stopped, "stopped flag must be set")
+	mgr.accountsLock.Unlock()
+
+	// Post-stop disconnect must be ignored.
+	mgr.OnPeerDisconnected(context.Background(), &nbpeer.Peer{
+		ID: "p2", AccountID: "acc-1", Ephemeral: true,
+		Status: &nbpeer.PeerStatus{Connected: false, LastSeen: timeNow()},
+	})
+	mgr.accountsLock.Lock()
+	require.Empty(t, mgr.accounts, "disconnects after Stop must be ignored")
+	mgr.accountsLock.Unlock()
+}
+
+// TestOnPeerConnected_IsNoop: the OnPeerConnected hook is preserved on
+// the interface but does nothing in the per-account model — the sweep
+// query filters connected peers at the DB level.
+func TestOnPeerConnected_IsNoop(t *testing.T) {
+	mockStore := &MockStore{account: newAccountWithId(context.Background(), "acc-1", "", "", false)}
+	withFakeClock(t, time.Now())
+
+	ctrl := gomock.NewController(t)
+	peersMgr := peers.NewMockManager(ctrl)
+
+	mgr := newManagerForTest(t, mockStore, peersMgr)
+	mgr.OnPeerDisconnected(context.Background(), &nbpeer.Peer{
+		ID: "p1", AccountID: "acc-1", Ephemeral: true,
+		Status: &nbpeer.PeerStatus{Connected: false, LastSeen: timeNow()},
+	})
+	mgr.accountsLock.Lock()
+	require.Len(t, mgr.accounts, 1, "disconnect should track the account")
+	mgr.accountsLock.Unlock()
+
+	mgr.OnPeerConnected(context.Background(), &nbpeer.Peer{ID: "p1", AccountID: "acc-1", Ephemeral: true})
+	mgr.accountsLock.Lock()
+	require.Len(t, mgr.accounts, 1, "OnPeerConnected must be a no-op")
+	mgr.accountsLock.Unlock()
+}
+
+// TestSweep_StoreErrorReArms: if the stale-peer query fails, the
+// account must remain tracked and a follow-up sweep gets scheduled.
+func TestSweep_StoreErrorReArms(t *testing.T) {
+	mockStore := &erroringStore{
+		MockStore: MockStore{account: newAccountWithId(context.Background(), "acc-1", "", "", false)},
+	}
+	getNow, setNow := withFakeClock(t, time.Now())
+
+	ctrl := gomock.NewController(t)
+	peersMgr := peers.NewMockManager(ctrl)
+
+	mgr := newManagerForTest(t, mockStore, peersMgr)
+
+	p := &nbpeer.Peer{ID: "p1", AccountID: "acc-1", Ephemeral: true,
+		Status: &nbpeer.PeerStatus{Connected: false, LastSeen: getNow()}}
+	mockStore.account.Peers[p.ID] = p
+	mgr.OnPeerDisconnected(context.Background(), p)
+
+	mockStore.fail.Store(true)
+	setNow(getNow().Add(mgr.lifeTime + 5*mgr.cleanupWindow))
+
+	// Wait until the failing sweep has run at least once.
+	require.Eventually(t, func() bool { return mockStore.failedCalls.Load() >= 1 },
+		2*time.Second, 5*time.Millisecond, "expected at least one failing sweep")
+
+	mgr.accountsLock.Lock()
+	_, stillTracked := mgr.accounts["acc-1"]
+	mgr.accountsLock.Unlock()
+	require.True(t, stillTracked, "account must remain tracked after a sweep error")
+
+	// Recover and ensure the rearmed sweep cleans up.
+	peersMgr.EXPECT().
+		DeletePeers(gomock.Any(), "acc-1", gomock.Any(), gomock.Any(), true).
+		DoAndReturn(func(_ context.Context, _ string, peerIDs []string, _ string, _ bool) error {
+			mockStore.mu.Lock()
+			for _, id := range peerIDs {
+				delete(mockStore.account.Peers, id)
+			}
+			mockStore.mu.Unlock()
+			return nil
+		}).AnyTimes()
+	mockStore.fail.Store(false)
+
+	require.Eventually(t, func() bool {
+		mockStore.mu.Lock()
+		defer mockStore.mu.Unlock()
+		_, gone := mockStore.account.Peers["p1"]
+		return !gone
+	}, 2*time.Second, 5*time.Millisecond, "rearmed sweep should clean up after the store recovers")
+}
+
+// erroringStore is a MockStore that can be flipped into a failing mode
+// to exercise the sweep's error-rearm path.
+type erroringStore struct {
+	MockStore
+	fail        atomic.Bool
+	failedCalls atomic.Int32
+}
+
+func (s *erroringStore) GetStaleEphemeralPeerIDsForAccount(ctx context.Context, accountID string, olderThan time.Time) ([]string, error) {
+	if s.fail.Load() {
+		s.failedCalls.Add(1)
+		return nil, errors.New("synthetic store error")
+	}
+	return s.MockStore.GetStaleEphemeralPeerIDsForAccount(ctx, accountID, olderThan)
+}
+
+// TestDefaultInitialLoadDelay confirms the jitter falls inside the
+// documented [8m, 10m) range — sanity check for the production timer.
+func TestDefaultInitialLoadDelay(t *testing.T) {
+	for i := 0; i < 1000; i++ {
+		d := defaultInitialLoadDelay()
+		assert.GreaterOrEqual(t, d, initialLoadMinDelay)
+		assert.Less(t, d, initialLoadMaxDelay)
 	}
 }
 
@@ -351,3 +596,7 @@ func newAccountWithId(ctx context.Context, accountID, userID, domain string, dis
 	}
 	return acc
 }
+
+// silence the import "github.com/netbirdio/netbird/management/internals/modules/peers/ephemeral"
+// (still needed indirectly for ephemeral.EphemeralLifeTime in production paths).
+var _ = ephemeral.EphemeralLifeTime
diff --git a/management/server/store/sql_store.go b/management/server/store/sql_store.go
index 8cf37de56..12838d4b8 100644
--- a/management/server/store/sql_store.go
+++ b/management/server/store/sql_store.go
@@ -3463,6 +3463,49 @@ func (s *SqlStore) GetAllEphemeralPeers(ctx context.Context, lockStrength Lockin
 	return allEphemeralPeers, nil
 }
 
+// GetStaleEphemeralPeerIDsForAccount returns IDs of disconnected
+// ephemeral peers in the given account whose last_seen is strictly
+// older than olderThan.
+func (s *SqlStore) GetStaleEphemeralPeerIDsForAccount(ctx context.Context, accountID string, olderThan time.Time) ([]string, error) {
+	var ids []string
+	err := s.db.WithContext(ctx).
+		Model(&nbpeer.Peer{}).
+		Where("account_id = ? AND ephemeral = ? AND peer_status_connected = ? AND peer_status_last_seen < ?",
+			accountID, true, false, olderThan).
+		Pluck("id", &ids).Error
+	if err != nil {
+		log.WithContext(ctx).Errorf("failed to query stale ephemeral peers for account %s: %v", accountID, err)
+		return nil, status.Errorf(status.Internal, "query stale ephemeral peers")
+	}
+	return ids, nil
+}
+
+// GetEphemeralAccountsLastDisconnect returns the latest peer_status_last_seen
+// per account across disconnected ephemeral peers. Returns one entry per
+// account that has at least one such peer.
+func (s *SqlStore) GetEphemeralAccountsLastDisconnect(ctx context.Context) (map[string]time.Time, error) {
+	type row struct {
+		AccountID string
+		LastSeen  time.Time
+	}
+	var rows []row
+	err := s.db.WithContext(ctx).
+		Model(&nbpeer.Peer{}).
+		Select("account_id, MAX(peer_status_last_seen) AS last_seen").
+		Where("ephemeral = ? AND peer_status_connected = ?", true, false).
+		Group("account_id").
+		Scan(&rows).Error
+	if err != nil {
+		log.WithContext(ctx).Errorf("failed to load ephemeral-account last disconnect map: %v", err)
+		return nil, status.Errorf(status.Internal, "load ephemeral accounts")
+	}
+	out := make(map[string]time.Time, len(rows))
+	for _, r := range rows {
+		out[r.AccountID] = r.LastSeen
+	}
+	return out, nil
+}
+
 // DeletePeer removes a peer from the store.
 func (s *SqlStore) DeletePeer(ctx context.Context, accountID string, peerID string) error {
 	result := s.db.Delete(&nbpeer.Peer{}, accountAndIDQueryCondition, accountID, peerID)
diff --git a/management/server/store/store.go b/management/server/store/store.go
index a723c1fc3..b5e6b3ed8 100644
--- a/management/server/store/store.go
+++ b/management/server/store/store.go
@@ -165,6 +165,15 @@ type Store interface {
 	GetAccountPeersWithExpiration(ctx context.Context, lockStrength LockingStrength, accountID string) ([]*nbpeer.Peer, error)
 	GetAccountPeersWithInactivity(ctx context.Context, lockStrength LockingStrength, accountID string) ([]*nbpeer.Peer, error)
 	GetAllEphemeralPeers(ctx context.Context, lockStrength LockingStrength) ([]*nbpeer.Peer, error)
+	// GetStaleEphemeralPeerIDsForAccount returns the IDs of disconnected
+	// ephemeral peers whose last_seen is strictly older than olderThan,
+	// scoped to a single account. Used by the per-account cleanup sweep.
+	GetStaleEphemeralPeerIDsForAccount(ctx context.Context, accountID string, olderThan time.Time) ([]string, error)
+	// GetEphemeralAccountsLastDisconnect returns, for every account that
+	// has at least one disconnected ephemeral peer, the most recent
+	// last_seen across that account's disconnected ephemeral peers. Used
+	// to reconstruct the per-account cleanup tracker after a restart.
+	GetEphemeralAccountsLastDisconnect(ctx context.Context) (map[string]time.Time, error)
 	SavePeer(ctx context.Context, accountID string, peer *nbpeer.Peer) error
 	SavePeerStatus(ctx context.Context, accountID, peerID string, status nbpeer.PeerStatus) error
 	// MarkPeerConnectedIfNewerSession sets the peer to connected with the
diff --git a/management/server/store/store_mock.go b/management/server/store/store_mock.go
index d51629606..d9ab09f8a 100644
--- a/management/server/store/store_mock.go
+++ b/management/server/store/store_mock.go
@@ -1376,6 +1376,36 @@ func (mr *MockStoreMockRecorder) GetAllEphemeralPeers(ctx, lockStrength interfac
 	return mr.mock.ctrl.RecordCallWithMethodType(mr.mock, "GetAllEphemeralPeers", reflect.TypeOf((*MockStore)(nil).GetAllEphemeralPeers), ctx, lockStrength)
 }
 
+// GetStaleEphemeralPeerIDsForAccount mocks base method.
+func (m *MockStore) GetStaleEphemeralPeerIDsForAccount(ctx context.Context, accountID string, olderThan time.Time) ([]string, error) {
+	m.ctrl.T.Helper()
+	ret := m.ctrl.Call(m, "GetStaleEphemeralPeerIDsForAccount", ctx, accountID, olderThan)
+	ret0, _ := ret[0].([]string)
+	ret1, _ := ret[1].(error)
+	return ret0, ret1
+}
+
+// GetStaleEphemeralPeerIDsForAccount indicates an expected call of GetStaleEphemeralPeerIDsForAccount.
+func (mr *MockStoreMockRecorder) GetStaleEphemeralPeerIDsForAccount(ctx, accountID, olderThan interface{}) *gomock.Call {
+	mr.mock.ctrl.T.Helper()
+	return mr.mock.ctrl.RecordCallWithMethodType(mr.mock, "GetStaleEphemeralPeerIDsForAccount", reflect.TypeOf((*MockStore)(nil).GetStaleEphemeralPeerIDsForAccount), ctx, accountID, olderThan)
+}
+
+// GetEphemeralAccountsLastDisconnect mocks base method.
+func (m *MockStore) GetEphemeralAccountsLastDisconnect(ctx context.Context) (map[string]time.Time, error) {
+	m.ctrl.T.Helper()
+	ret := m.ctrl.Call(m, "GetEphemeralAccountsLastDisconnect", ctx)
+	ret0, _ := ret[0].(map[string]time.Time)
+	ret1, _ := ret[1].(error)
+	return ret0, ret1
+}
+
+// GetEphemeralAccountsLastDisconnect indicates an expected call of GetEphemeralAccountsLastDisconnect.
+func (mr *MockStoreMockRecorder) GetEphemeralAccountsLastDisconnect(ctx interface{}) *gomock.Call {
+	mr.mock.ctrl.T.Helper()
+	return mr.mock.ctrl.RecordCallWithMethodType(mr.mock, "GetEphemeralAccountsLastDisconnect", reflect.TypeOf((*MockStore)(nil).GetEphemeralAccountsLastDisconnect), ctx)
+}
+
 // GetAllProxyAccessTokens mocks base method.
 func (m *MockStore) GetAllProxyAccessTokens(ctx context.Context, lockStrength LockingStrength) ([]*types2.ProxyAccessToken, error) {
 	m.ctrl.T.Helper()
diff --git a/management/server/telemetry/ephemeral_metrics.go b/management/server/telemetry/ephemeral_metrics.go
index a7fb432f8..3424366ca 100644
--- a/management/server/telemetry/ephemeral_metrics.go
+++ b/management/server/telemetry/ephemeral_metrics.go
@@ -7,9 +7,9 @@ import (
 )
 
 // EphemeralPeersMetrics tracks the ephemeral peer cleanup pipeline: how
-// many peers are currently scheduled for deletion, how many tick runs
-// the cleaner has performed, how many peers it has removed, and how
-// many delete batches failed.
+// many accounts are currently being tracked for cleanup, how many sweep
+// runs deleted at least one peer, how many peers have been removed, and
+// how many delete batches failed.
 type EphemeralPeersMetrics struct {
 	ctx context.Context
 
@@ -21,16 +21,16 @@ type EphemeralPeersMetrics struct {
 
 // NewEphemeralPeersMetrics constructs the ephemeral cleanup counters.
 func NewEphemeralPeersMetrics(ctx context.Context, meter metric.Meter) (*EphemeralPeersMetrics, error) {
-	pending, err := meter.Int64UpDownCounter("management.ephemeral.peers.pending",
+	pending, err := meter.Int64UpDownCounter("management.ephemeral.accounts.tracked",
 		metric.WithUnit("1"),
-		metric.WithDescription("Number of ephemeral peers currently waiting to be cleaned up"))
+		metric.WithDescription("Number of accounts currently tracked for ephemeral peer cleanup"))
 	if err != nil {
 		return nil, err
 	}
 
 	cleanupRuns, err := meter.Int64Counter("management.ephemeral.cleanup.runs.counter",
 		metric.WithUnit("1"),
-		metric.WithDescription("Number of ephemeral cleanup ticks that processed at least one peer"))
+		metric.WithDescription("Number of ephemeral cleanup sweeps that deleted at least one peer"))
 	if err != nil {
 		return nil, err
 	}
@@ -61,7 +61,8 @@ func NewEphemeralPeersMetrics(ctx context.Context, meter metric.Meter) (*Ephemer
 // All methods are nil-receiver safe so callers that haven't wired metrics
 // (tests, self-hosted with metrics off) can invoke them unconditionally.
 
-// IncPending bumps the pending gauge when a peer is added to the cleanup list.
+// IncPending bumps the tracked-accounts gauge when a new account
+// becomes eligible for ephemeral cleanup tracking.
 func (m *EphemeralPeersMetrics) IncPending() {
 	if m == nil {
 		return
@@ -69,8 +70,8 @@ func (m *EphemeralPeersMetrics) IncPending() {
 	m.pending.Add(m.ctx, 1)
 }
 
-// AddPending bumps the pending gauge by n — used at startup when the
-// initial set of ephemeral peers is loaded from the store.
+// AddPending bumps the tracked-accounts gauge by n — used at startup
+// when the catch-up query seeds the tracker.
 func (m *EphemeralPeersMetrics) AddPending(n int64) {
 	if m == nil || n <= 0 {
 		return
@@ -78,9 +79,8 @@ func (m *EphemeralPeersMetrics) AddPending(n int64) {
 	m.pending.Add(m.ctx, n)
 }
 
-// DecPending decreases the pending gauge — used both when a peer reconnects
-// before its deadline (removed from the list) and when a cleanup tick
-// actually deletes it.
+// DecPending decreases the tracked-accounts gauge when an account is
+// dropped from the tracker (no more disconnects to chase).
 func (m *EphemeralPeersMetrics) DecPending(n int64) {
 	if m == nil || n <= 0 {
 		return