Skip to content

"Vibe-Trading: Your Personal Trading Agent"

fix(memory): the compression pass skips a file that is gone, not an id that was acted on (#1450)

Haozhe Wu451c554aa1ec
diff --git a/agent/src/memory/lifecycle.py b/agent/src/memory/lifecycle.py
index 417a42049..327771316 100644
--- a/agent/src/memory/lifecycle.py
+++ b/agent/src/memory/lifecycle.py
@@ -205,7 +205,6 @@ class MemoryLifecycle:
now = time.time()
actions: list[dict] = []
- gc_acted_ids: set[str] = set()
for entry in entries:
age_days = (now - entry.created_at) / 86400.0
@@ -238,17 +237,17 @@ class MemoryLifecycle:
# Tier 1: force archive even if classified as delete
effective = "archive" if not self.ENABLE_DELETE else action
self._execute_gc_action(entry, effective)
- gc_acted_ids.add(entry.id)
self._append_gc_log(actions, dry_run)
# Tier 2: Trigger compression for aged entries. Skipped on dry_run:
# compression rewrites entry bodies and frontmatter in place, and a
- # dry run must not mutate the files it is only auditing. Entries
- # already archived/deleted above are skipped too: _execute_gc_action
- # already renamed or unlinked entry.path on disk, so re-using this
- # same pre-Tier-1 snapshot for them would try to compress a file
- # that no longer exists at that location.
+ # dry run must not mutate the files it is only auditing. An entry
+ # Tier 1 archived or deleted is skipped too: _execute_gc_action renamed
+ # or unlinked entry.path, and this is the pre-Tier-1 snapshot. The
+ # check is on the path, not on entry.id: ids are a six-hex hash two
+ # files can share, and an archive step that returned early (lock
+ # timeout, OSError) leaves its file in place to be compressed (#1450).
from src.config.accessor import get_env_config
if not dry_run and get_env_config().memory.compression_enabled:
try:
@@ -256,7 +255,7 @@ class MemoryLifecycle:
pipeline = CompressionPipeline(self._memory._dir)
now_ts = time.time()
for entry in entries:
- if entry.id in gc_acted_ids:
+ if not entry.path.exists():
continue
target = pipeline.should_compress(
compression_level=entry.compression_level,
diff --git a/agent/tests/test_memory_gc.py b/agent/tests/test_memory_gc.py
index 441ed0a0c..8c7bfb17b 100644
--- a/agent/tests/test_memory_gc.py
+++ b/agent/tests/test_memory_gc.py
@@ -260,6 +260,51 @@ class TestRunGC:
assert "Cannot archive non-existent file" not in caplog.text
assert "_write_compressed" not in caplog.text
+ def test_gc_compresses_a_healthy_entry_that_shares_an_archived_entrys_id(
+ self, tmp_path: Path, monkeypatch
+ ) -> None:
+ """The Tier 2 skip is keyed on the file, not on the entry id (#1450).
+
+ Ids are a six-hex hash (``sha256(...)[:6]`` when the frontmatter has
+ none), so two files can carry the same one. Skipping by id would leave
+ the healthy file uncompressed because its twin was archived.
+ """
+ monkeypatch.setenv("VT_MEMORY_GC", "1")
+ monkeypatch.setenv("VT_MEMORY_DECAY", "1")
+ monkeypatch.setenv("VT_MEMORY_COMPRESSION", "1")
+ now = time.time()
+ old_iso = time.strftime("%Y-%m-%dT%H:%M:%S", time.gmtime(now - 30 * 86400))
+ _create_memory_file(
+ tmp_path,
+ "old-and-stale",
+ quality_score=0.01,
+ access_count=0,
+ created_at=old_iso,
+ last_accessed=old_iso,
+ entry_id="ab12cd",
+ )
+ healthy = _create_memory_file(
+ tmp_path,
+ "young-but-unread",
+ content=" ".join(
+ f"Sentence number {i} discusses trading topic {i} in detail." for i in range(12)
+ ),
+ keywords=["alpha", "momentum"],
+ created_at=time.strftime("%Y-%m-%dT%H:%M:%S", time.gmtime(now - 2 * 86400)),
+ last_accessed=time.strftime("%Y-%m-%dT%H:%M:%S", time.gmtime(now - 9 * 86400)),
+ entry_id="ab12cd",
+ )
+ before = healthy.read_text(encoding="utf-8")
+ lc = MemoryLifecycle(PersistentMemory(memory_dir=tmp_path))
+
+ actions = lc.run_gc(dry_run=False)
+
+ assert [a["action"] for a in actions] == ["archive"]
+ assert (tmp_path / "archive" / "project_old-and-stale.md").exists()
+ after = healthy.read_text(encoding="utf-8")
+ assert after != before, "the healthy twin was skipped because it shares the archived id"
+ assert "compression_level: daily" in after
+
# ---------------------------------------------------------------------------
# 2. find_relevant importance weighting