[Bug]: Knowledge-base watcher stores sync time instead of source mtime

[BunnyTeddy]

Problem

The automatic MemorySyncManager path for files under knowledge_base/ calls LongTermMemoryStore.index_file() without the file's source mtime. index_file() therefore defaults the database files.mtime value to the current sync time. The direct memory ingest path already passes the document mtime, so watcher-indexed documents behave inconsistently.

This makes file freshness metadata inaccurate and can distort retrieval recency scoring for older documents.

Reproduction

1. Create knowledge_base/old.md and set its filesystem mtime to a timestamp well in the past.
2. Start MemorySyncManager (or wait for its watcher to index the file).
3. Read the stored mtime with LongTermMemoryStore.get_file_mtimes(["knowledge_base/old.md"]).
4. Observe that it is close to the sync time rather than the source file's mtime.

Expected behavior

Automatic knowledge-base synchronization passes the source file mtime to index_file(), matching ingest_document(), so the persisted metadata reflects the document on disk.

Acceptance criteria

• Initial and watcher-driven knowledge-base indexing pass the file's filesystem mtime to index_file().
• A regression test proves the captured mtime equals the source file mtime.
• Existing memory-file sync, public APIs, persistence schema, and migrations remain unchanged.
• No network or provider dependency is required by the test.

[BunnyTeddy]

I would like to work on this. I reproduced the watcher path dropping the source mtime and will submit a focused fix plus an offline regression test, without changing the public memory API.

[BunnyTeddy]

Contributor access is read-only for label mutations, so I could not apply the requested labels. Maintainer action needed: type: bug, area: memory, priority: p2.

[andreapn]

Accepted, p3. Confirmed. _do_file_sync calls:

n = await self._store.index_file(path=rel_path, content=content, source=source)

index_file accepts mtime: float | None = None (src/agentos/memory/backend.py:30) and ingest.py:218 passes p.stat().st_mtime on the manual path — so the inconsistency between memory ingest and watcher indexing is real.

Convenient detail for whoever picks this up: _scan_files() has already called path.stat().st_mtime and the value is sitting in the current dict, so the fix does not need an extra stat call.

One scope note: the issue frames this as knowledge-base-only, but the index_file call is shared by both the knowledge_base/ and MEMORY.md branches. Pass the source mtime for both — there is no reason for the memory branch to keep recording sync time.

PR #650 covers this.