diff --git a/.agent-tasks/2026-03-20--4--log-accumulator-crash/discarded-threads.md b/.agent-tasks/2026-03-20--4--log-accumulator-crash/discarded-threads.md
new file mode 100644
index 0000000..8ca433c
--- /dev/null
+++ b/.agent-tasks/2026-03-20--4--log-accumulator-crash/discarded-threads.md
@@ -0,0 +1,3 @@
+# Discarded threads
+
+(none)
diff --git a/.agent-tasks/2026-03-20--4--log-accumulator-crash/open-questions.md b/.agent-tasks/2026-03-20--4--log-accumulator-crash/open-questions.md
new file mode 100644
index 0000000..5b5439d
--- /dev/null
+++ b/.agent-tasks/2026-03-20--4--log-accumulator-crash/open-questions.md
@@ -0,0 +1,3 @@
+# Open questions
+
+(none)
diff --git a/.agent-tasks/2026-03-20--4--log-accumulator-crash/plan.md b/.agent-tasks/2026-03-20--4--log-accumulator-crash/plan.md
new file mode 100644
index 0000000..a562d2e
--- /dev/null
+++ b/.agent-tasks/2026-03-20--4--log-accumulator-crash/plan.md
@@ -0,0 +1,36 @@
+# Implementation Plan
+
+## Step 1: Fix task result handler to properly clean up
+
+In `handle_info({ref, result}, state)` (line 124-134):
+- Call `Process.demonitor(ref, [:flush])` to prevent the `:DOWN` message from arriving
+- Actually remove the ref from `pending_tasks` (the current code has a comment but doesn't do it)
+
+## Step 2: Fix `block_until_any_task_ready` to demonitor on result receipt
+
+When the `receive` block in `block_until_any_task_ready` catches `{ref, _result}`:
+- Call `Process.demonitor(ref, [:flush])` to prevent orphaned `:DOWN` messages
+
+## Step 3: Add catch-all `handle_info` clause
+
+Add a catch-all at the bottom of the `handle_info` clauses that silently ignores any unexpected messages. This is a safety net for any edge cases we haven't anticipated.
+
+## Step 4: Add test for the crash scenario
+
+Write a test that:
+- Sets up a LogAccumulator with concurrent request limit of 1
+- Sends enough logs to trigger `block_until_any_task_ready`
+- Makes the export endpoint unavailable so tasks fail
+- Verifies the LogAccumulator doesn't crash
+
+---
+
+## Review discussion
+
+**Reviewer concern:** Adding a catch-all `handle_info` could hide future bugs.
+
+**Response:** Since this module uses `@behaviour GenServer` and is used as a `logger_olp` callback, unexpected messages are a real possibility (e.g., `:EXIT` from `trap_exit`, or stray messages from the OTP logger framework). The catch-all only applies after all specific handlers have been tried. The `:DOWN` and task result message types are well-understood, and the catch-all acts as a safety net.
+
+**Reviewer concern:** Should we log stray messages?
+
+**Response:** No — logging from inside a log handler risks recursion. Silent discard is the correct behavior here.
diff --git a/.agent-tasks/2026-03-20--4--log-accumulator-crash/progress.md b/.agent-tasks/2026-03-20--4--log-accumulator-crash/progress.md
new file mode 100644
index 0000000..440b9b8
--- /dev/null
+++ b/.agent-tasks/2026-03-20--4--log-accumulator-crash/progress.md
@@ -0,0 +1,15 @@
+# Progress log
+
+## Timeline
+
+- **2026-03-20 Start**: Claimed task, read issue #18 and LogAccumulator source code
+- **2026-03-20**: Identified root cause — race between `block_until_any_task_ready` consuming `{ref, result}` and subsequent `:DOWN` arriving via `handle_info` with ref already removed from `pending_tasks`
+- **2026-03-20**: Also found secondary bug — `handle_info({ref, result}, state)` doesn't actually remove ref from `pending_tasks` despite the comment
+- **2026-03-20**: Implemented fix with 3 changes: proper demonitor+cleanup in handle_info, demonitor+flush in block_until_any_task_ready, and catch-all handle_info clause
+- **2026-03-20**: Added integration test, all 45 tests pass
+- **2026-03-20**: Opened PR https://github.com/electric-sql/elixir-otel-metric-exporter/pull/31
+
+## Operational issues
+
+- Had to run `mix deps.get` to download `protobuf` dependency before tests would run
+- Test design required careful handling of the feedback loop: failed exports generate Logger.debug logs which themselves get queued as log events, causing cascading export attempts. Used `retry: false` config and `Bypass.stub` (not `expect_once`) to handle this gracefully
diff --git a/.agent-tasks/2026-03-20--4--log-accumulator-crash/prompt.md b/.agent-tasks/2026-03-20--4--log-accumulator-crash/prompt.md
new file mode 100644
index 0000000..5d5308e
--- /dev/null
+++ b/.agent-tasks/2026-03-20--4--log-accumulator-crash/prompt.md
@@ -0,0 +1,8 @@
+# Task prompt
+
+Assigned issue: electric-sql/alco-agent-tasks#4
+Upstream issue: electric-sql/elixir-otel-metric-exporter#18
+
+**Title:** When export fails repeatedly due to the remote server being unavailable, LogAccumulator crashes hard
+
+**Description:** Investigate and implement a fix for the LogAccumulator crashing with `FunctionClauseError` when it receives `:DOWN` messages from export tasks that it no longer tracks in `pending_tasks`.
diff --git a/.agent-tasks/2026-03-20--4--log-accumulator-crash/task.md b/.agent-tasks/2026-03-20--4--log-accumulator-crash/task.md
new file mode 100644
index 0000000..b2e4c82
--- /dev/null
+++ b/.agent-tasks/2026-03-20--4--log-accumulator-crash/task.md
@@ -0,0 +1,27 @@
+# Task: Fix LogAccumulator crash on unmatched :DOWN messages
+
+## Problem
+
+The `LogAccumulator` GenServer (used as a `logger_olp` callback) crashes with a `FunctionClauseError` when it receives a `:DOWN` message whose ref is not present in `pending_tasks`.
+
+## Root Cause
+
+There are two code paths that handle task completion messages:
+
+1. **`handle_info/2`** — processes messages delivered via the normal GenServer message loop
+2. **`block_until_any_task_ready/1`** — uses a raw `receive` to consume messages from the mailbox when the process is blocking
+
+The race condition:
+- `block_until_any_task_ready` receives `{ref, result}` from a completed task and removes `ref` from `pending_tasks`
+- The subsequent `:DOWN` message for that same task process arrives via `handle_info`
+- The guard `is_map_key(state.pending_tasks, ref)` fails on the `:DOWN` handler
+- No catch-all clause exists → `FunctionClauseError` crash
+
+Secondary bug: `handle_info({ref, result}, state)` at line 124 says "Remove the task from the pending tasks map" but the returned state is unchanged — the ref is never actually removed.
+
+## Fix Strategy
+
+1. In `handle_info({ref, result}, state)`: demonitor with flush and actually remove ref from `pending_tasks`
+2. In `block_until_any_task_ready`: when receiving `{ref, result}`, demonitor with flush to prevent orphaned `:DOWN`
+3. Add a catch-all `handle_info` clause to silently ignore any stray `:DOWN` or task result messages
+4. Add a test that exercises the crash scenario
diff --git a/lib/otel_metric_exporter/log_accumulator.ex b/lib/otel_metric_exporter/log_accumulator.ex
index db8cc19..b6f4f63 100644
--- a/lib/otel_metric_exporter/log_accumulator.ex
+++ b/lib/otel_metric_exporter/log_accumulator.ex
@@ -123,22 +123,28 @@ defmodule OtelMetricExporter.LogAccumulator do
 
   def handle_info({ref, result}, state)
       when is_map_key(state.pending_tasks, ref) do
+    Process.demonitor(ref, [:flush])
+
     if match?({:error, _}, result) do
       Logger.debug(
         "Error sending logs to #{state.api.config.otlp_endpoint}: #{inspect(elem(result, 1))}"
       )
     end
 
-    # Remove the task from the pending tasks map
-    {:noreply, state}
+    {:noreply, %{state | pending_tasks: Map.delete(state.pending_tasks, ref)}}
   end
 
   def handle_info({:DOWN, ref, :process, _, _}, state)
       when is_map_key(state.pending_tasks, ref) do
-    # Remove the task from the pending tasks map
     {:noreply, %{state | pending_tasks: Map.delete(state.pending_tasks, ref)}}
   end
 
+  # Catch-all for stray messages (e.g. late :DOWN or task results after ref was
+  # already cleaned up by block_until_any_task_ready or a prior handler).
+  def handle_info(_msg, state) do
+    {:noreply, state}
+  end
+
   def terminate(_reason, state) do
     # Send any remaining logs if possible
     send_events_via_task(state)
@@ -179,11 +185,10 @@ defmodule OtelMetricExporter.LogAccumulator do
     # from a task that we started
     receive do
       {ref, _result} when is_map_key(pending_tasks, ref) ->
-        # Remove the task from the pending tasks map
+        Process.demonitor(ref, [:flush])
         %{state | pending_tasks: Map.delete(pending_tasks, ref)}
 
       {:DOWN, ref, :process, _, _} when is_map_key(pending_tasks, ref) ->
-        # Remove the task from the pending tasks map
         %{state | pending_tasks: Map.delete(pending_tasks, ref)}
     end
   end
diff --git a/test/otel_metric_exporter/log_handler_integration_test.exs b/test/otel_metric_exporter/log_handler_integration_test.exs
index 6a6f966..4a3021d 100644
--- a/test/otel_metric_exporter/log_handler_integration_test.exs
+++ b/test/otel_metric_exporter/log_handler_integration_test.exs
@@ -444,4 +444,37 @@ defmodule OtelMetricExporter.LogHandlerIntegrationTest do
              }
            ] = logs
   end
+
+  test "LogAccumulator survives repeated export failures", %{
+    bypass: bypass,
+    handler_id: handler_id,
+    config: initial_config
+  } do
+    # Configure with small buffer, single concurrent request, and no retries.
+    # This forces block_until_any_task_ready to be exercised, which is the
+    # code path that previously caused FunctionClauseError on stray :DOWN messages.
+    new_config = %{max_buffer_size: 1, debounce_ms: 5, otlp_concurrent_requests: 1, retry: false}
+    merged_config = Map.merge(initial_config, new_config)
+    :ok = :logger.set_handler_config(handler_id, %{config: merged_config})
+
+    # Return 500 for all requests so exports fail immediately (no retry)
+    Bypass.stub(bypass, "POST", "/v1/logs", fn conn ->
+      Plug.Conn.resp(conn, 500, "")
+    end)
+
+    # Send a burst of logs to trigger multiple export attempts.
+    # With max_buffer_size=1 and concurrent_requests=1, this will force
+    # the accumulator into block_until_any_task_ready repeatedly.
+    for i <- 1..5 do
+      Logger.info("failure-test-#{i}")
+      Process.sleep(10)
+    end
+
+    # Wait for tasks to complete
+    Process.sleep(200)
+
+    # Verify the handler is still alive by checking it's still registered
+    handlers = :logger.get_handler_ids()
+    assert handler_id in handlers, "LogAccumulator should still be alive after export failures"
+  end
 end