diff --git a/tines-ros2/src/node.rs b/tines-ros2/src/node.rs
index 38fc030a..ec571a24 100644
--- a/tines-ros2/src/node.rs
+++ b/tines-ros2/src/node.rs
@@ -175,6 +175,11 @@ impl TinesNode {
     /// One bridge tick: consume inbound ROS2 observations, drive the tines
     /// exchange, publish live topics, and publish any confidence events.
     ///
+    /// `now` is a monotonic millisecond [`Tick`]: milliseconds since the host
+    /// process started (a single `Instant`-based clock, `0` at startup), never
+    /// wall-clock time. Pass the same clock's value every call so ages and
+    /// addressed lifetimes advance in real milliseconds.
+    ///
     /// Returns the number of *new* observations stored this round (from
     /// [`TinesNodeCore::poll`]).
     pub fn poll(&mut self, now: Tick) -> Result<usize, TinesNodeError> {
diff --git a/tines/src/clock.rs b/tines/src/clock.rs
new file mode 100644
index 00000000..c59deab7
--- /dev/null
+++ b/tines/src/clock.rs
@@ -0,0 +1,94 @@
+//! The host's single monotonic millisecond clock.
+//!
+//! [`HostClock`] is a single [`std::time::Instant`] captured at process start.
+//! Its [`now`](HostClock::now) returns the milliseconds elapsed since that
+//! capture as a [`u64`] tines `Tick`.
+//!
+//! It is monotonic, so an NTP step (forward or backwards) can't rewind it —
+//! ages and addressed lifetimes keep advancing instead of saturating or
+//! jumping. And it is expressed in milliseconds, the unit the tines `Tick`
+//! expects: an observation with `lifetime_ms = 1000` expires after ~1 s of the
+//! host's own ticks, not after ~1000 s (which is what a wall-clock *seconds*
+//! clock, the old `now_ms`, produced).
+//!
+//! There is one clock per process. `LogStore` is opened with `clock.now()`
+//! (which is `0` at startup, correct by construction), and the service uses
+//! the same clock for every `observe`/`belief` tick, so the store and the
+//! service never disagree about what "now" is.
+//!
+//! Wall-clock time (`SystemTime`, milliseconds since the Unix epoch) is used
+//! only for `observed_at`, and only when the host is configured to trust its
+//! own clock — never for the monotonic tick.
+
+use std::time::Instant;
+
+/// One process-wide monotonic millisecond clock.
+///
+/// Copy cheaply (a single [`Instant`]); pass it to whoever needs "now".
+#[derive(Debug, Clone)]
+pub struct HostClock {
+    /// The instant this clock measures elapsed time from (process start).
+    start: Instant,
+}
+
+impl HostClock {
+    /// Capture the clock at process start.
+    ///
+    /// `now()` is `0` immediately after this, and grows monotonically.
+    pub fn start() -> Self {
+        Self {
+            start: Instant::now(),
+        }
+    }
+
+    /// Milliseconds elapsed since [`HostClock::start`].
+    ///
+    /// `0` right after [`HostClock::start`]; monotonic thereafter.
+    pub fn now(&self) -> u64 {
+        self.start.elapsed().as_millis() as u64
+    }
+}
+
+impl Default for HostClock {
+    fn default() -> Self {
+        Self::start()
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use std::time::Duration;
+
+    use super::HostClock;
+
+    /// The tines `Tick` is milliseconds, not seconds. Over a short real sleep
+    /// the clock must advance by the size of that sleep (hundreds of ms),
+    /// never by single digits as a wall-clock *seconds* clock would. Fast and
+    /// non-flaky: a generous lower bound absorbs CI jitter without a long sleep.
+    #[test]
+    fn now_advances_in_milliseconds_not_seconds() {
+        let clock = HostClock::start();
+        let first = clock.now();
+        std::thread::sleep(Duration::from_millis(150));
+        let second = clock.now();
+        let delta = second.saturating_sub(first);
+        assert!(
+            delta >= 120,
+            "HostClock::now must advance in milliseconds: expected >=120 ms over a 150 ms sleep, got {delta} ms",
+        );
+        // And it must be ms since process start, not epoch seconds: no process
+        // has lived a century, so the absolute value stays well under 60 s of
+        // millis. This would catch a regression back to `as_secs()`.
+        assert!(
+            second < 60_000,
+            "HostClock::now must be ms since process start, was {second}",
+        );
+    }
+
+    #[test]
+    fn now_is_zero_immediately_after_start() {
+        let clock = HostClock::start();
+        // No sleep: elapsed is a few microseconds, well under a millisecond.
+        assert!(clock.now() < 1, "clock starts at 0, was {}", clock.now());
+    }
+}
diff --git a/tines/src/ipc.rs b/tines/src/ipc.rs
index 940e55a0..f254b360 100644
--- a/tines/src/ipc.rs
+++ b/tines/src/ipc.rs
@@ -26,9 +26,11 @@
 //!
 //! # Sequence numbers
 //!
-//! `now` is the tines tick passed to `observe`/`belief`; the service reads it
-//! from its own wall clock (`now_ms()`) and ignores any caller-supplied value,
-//! so the wire request carries no timestamp.
+//! `now` is the tines tick passed to `observe`/`belief`. The service reads it
+//! from its own monotonic millisecond [`HostClock`] (an `Instant` captured at
+//! process start; see [`crate::clock`]) and ignores any caller-supplied value,
+//! so the wire request carries no timestamp. The store is opened with the same
+//! clock's `now` (0 at startup), so the service and the store agree on "now".
 
 use std::fmt::Write as _;
 use std::io::{Read, Write};
@@ -38,6 +40,8 @@ use std::sync::{Arc, Mutex};
 
 use tines_core::{Node, Store, Transport};
 
+use crate::HostClock;
+
 const MAX_FRAME: usize = 1 << 20; // 1 MiB cap on a single JSON frame
 
 /// A confidence event handed out by the IPC `events` request.
@@ -183,9 +187,18 @@ where
 /// Serve IPC on `socket_path` until the listener dies or a response write
 /// fails. Blocks; each accepted connection is served request-by-request.
 ///
+/// `now` for each request comes from `clock`, the host's single monotonic
+/// millisecond [`HostClock`] (an `Instant` at process start), so the tick the
+/// exchange advances on is in real milliseconds. Open the store with the same
+/// clock's `now` (0 at startup) so the service and store agree on "now".
+///
 /// `Node<S, T>` may be any concrete store/transport, so the real binary passes
 /// `Node<LogStore, BtTransport>` and tests pass a mock.
-pub fn serve<S, T>(node: Node<S, T>, socket_path: impl AsRef<Path>) -> std::io::Result<()>
+pub fn serve<S, T>(
+    node: Node<S, T>,
+    socket_path: impl AsRef<Path>,
+    clock: HostClock,
+) -> std::io::Result<()>
 where
     S: Store + 'static,
     T: Transport + 'static,
@@ -204,7 +217,7 @@ where
         loop {
             match read_request(&mut stream) {
                 Ok(req) => {
-                    let now = now_ms();
+                    let now = clock.now();
                     let resp = {
                         // Recover from a poisoned lock so one panicking
                         // handler can't take down the daemon.
@@ -273,13 +286,6 @@ fn error(msg: impl std::fmt::Display) -> Response {
     }
 }
 
-fn now_ms() -> u64 {
-    std::time::SystemTime::now()
-        .duration_since(std::time::SystemTime::UNIX_EPOCH)
-        .map(|d| d.as_secs())
-        .unwrap_or(0)
-}
-
 fn encode_hex(bytes: &[u8]) -> String {
     let mut s = String::with_capacity(bytes.len() * 2);
     for b in bytes {
@@ -312,3 +318,25 @@ fn hex_val(c: u8) -> Result<u8, String> {
         other => Err(format!("invalid hex char: {other:02x}")),
     }
 }
+
+#[cfg(test)]
+mod tests {
+    use std::time::Duration;
+
+    // Regression for F17: the tines `Tick` is milliseconds. A wall-clock
+    // *seconds* clock (the old `now_ms`) advances by at most ~1 over a
+    // short interval, so an observation's `lifetime_ms` would expire ~1000×
+    // too slowly. This asserts the service clock advances in milliseconds.
+    #[test]
+    fn tick_advances_in_milliseconds() {
+        let clock = crate::HostClock::start();
+        let a = clock.now();
+        std::thread::sleep(Duration::from_millis(150));
+        let b = clock.now();
+        let delta = b.saturating_sub(a);
+        assert!(
+            delta >= 120,
+            "tines tick advances in milliseconds: expected >=120 ms over a 150 ms sleep, got {delta} ms",
+        );
+    }
+}
diff --git a/tines/src/lib.rs b/tines/src/lib.rs
index 607196ce..cda26177 100644
--- a/tines/src/lib.rs
+++ b/tines/src/lib.rs
@@ -10,6 +10,8 @@
 //! The IPC protocol is documented in [`ipc`].
 #![deny(missing_docs)]
 
+pub mod clock;
 pub mod ipc;
 
+pub use clock::HostClock;
 pub use ipc::{handle, serve, BeliefOut, EventOut, Request, Response};
diff --git a/tines/src/main.rs b/tines/src/main.rs
index 8cb47948..0acb4323 100644
--- a/tines/src/main.rs
+++ b/tines/src/main.rs
@@ -11,7 +11,7 @@
 use std::error::Error;
 use std::path::Path;
 
-use tines::serve;
+use tines::{serve, HostClock};
 use tines_bt::{BluerRadio, BtTransport};
 use tines_core::{Node, NodeId};
 use tines_store::LogStore;
@@ -41,7 +41,8 @@ fn main() -> Result<(), Box<dyn Error + Send + Sync>> {
     let peer = parse_node_id(&config.peer)?;
 
     // LogStore is the Store (S); BtTransport wraps it as the Transport (T).
-    let store = LogStore::new(Path::new(&config.log_path), 0, config.first_seq)
+    let clock = HostClock::start();
+    let store = LogStore::new(Path::new(&config.log_path), clock.now(), config.first_seq)
         .map_err(|e| format!("failed to open log store: {e:?}"))?;
     let radio = BluerRadio::new(me, peer);
     let transport = BtTransport::new(me, peer, radio);
@@ -52,7 +53,7 @@ fn main() -> Result<(), Box<dyn Error + Send + Sync>> {
         config.socket_path,
         node_hex(me)
     );
-    serve(node, Path::new(&config.socket_path))?;
+    serve(node, Path::new(&config.socket_path), clock)?;
     Ok(())
 }
 
diff --git a/tines/tests/ipc.rs b/tines/tests/ipc.rs
index cfc87917..1cc6f813 100644
--- a/tines/tests/ipc.rs
+++ b/tines/tests/ipc.rs
@@ -14,7 +14,7 @@ use std::time::Duration;
 
 use heapless::Vec as HeaplessVec;
 use std::fmt::Write as _;
-use tines::{handle, serve, Request, Response};
+use tines::{handle, serve, HostClock, Request, Response};
 use tines_core::{
     consts, Event, Frame, FrameBuf, Node, NodeId, ObservationId, SendError, Transport,
 };
@@ -220,7 +220,7 @@ fn observe_and_belief_over_socket() {
     let socket = dir.join("ipc.sock");
     let node = build_node([9; 16], &dir);
     let serve_path = socket.clone();
-    let _handle = thread::spawn(move || serve(node, serve_path).unwrap());
+    let _handle = thread::spawn(move || serve(node, serve_path, HostClock::start()).unwrap());
 
     let mut stream = connect(&socket);
 
@@ -268,7 +268,7 @@ fn events_over_socket() {
     node.poll(0); // ack processed -> confidence event fires
 
     let serve_path = socket.clone();
-    let _handle = thread::spawn(move || serve(node, serve_path).unwrap());
+    let _handle = thread::spawn(move || serve(node, serve_path, HostClock::start()).unwrap());
     let mut stream = connect(&socket);
 
     rpc(&mut stream, &Request::Events);
@@ -322,7 +322,7 @@ fn malformed_frame_gets_error_and_connection_survives() {
     let socket = dir.join("ipc.sock");
     let node = build_node([3; 16], &dir);
     let serve_path = socket.clone();
-    let _handle = thread::spawn(move || serve(node, serve_path).unwrap());
+    let _handle = thread::spawn(move || serve(node, serve_path, HostClock::start()).unwrap());
     let mut stream = connect(&socket);
 
     // A length-prefixed frame whose body is not valid JSON.