hermes - 💡(How to fix) Fix OpenViking plugin: on_memory_write mirroring silently fails — 3 bugs [2 pull requests]

Bug Description

The OpenViking memory plugin (plugins/memory/openviking/__init__.py) has an on_memory_write hook that is supposed to mirror built-in memory() writes to the Viking knowledge base. In practice, the mirroring has never worked over 2-3 weeks of daily use. After investigating the source, three concrete bugs explain why.

Environment

• Hermes Agent v0.14.0 (2026.5.16), up to date
• OpenViking server running at localhost:1933, healthy
• memory.provider: openviking in config.yaml, hermes memory status shows "available ✓"
• WSL2 Ubuntu

Three bugs found in source

Bug 1: Only mirrors action="add" — replace/remove silently skipped

Line 642:

if not self._client or action != "add" or not content:
    return

The built-in memory() tool supports three actions: add, replace, remove. But on_memory_write only processes add. Every replace and remove call is dropped without any indication.

In practice, most long-lived memory entries are created once with add and then updated with replace multiple times. Those updates never reach Viking.

Bug 2: mode="create" causes collisions on re-add

Line 657:

client.post("/api/v1/content/write", {
    "uri": uri,
    "content": content,
    "mode": "create",
})

The URI is built with uuid.uuid4().hex[:12] — unique per call. But when the same memory entry is re-added (e.g. after a remove+add cycle, or when the built-in store is rebuilt), OpenViking rejects duplicate URIs with INVALID_ARGUMENT: resource is busy and cannot be written now. There is no mode="upsert" or mode="update" fallback.

Bug 3: All failures logged at logger.debug() — invisible in production

Line 660:

except Exception as e:
    logger.debug("OpenViking memory mirror failed: %s", e)

The mirror runs in a daemon thread and catches all exceptions at DEBUG level. With default log settings (INFO+), these failures produce zero visible output. Searching errors.log and agent.log for WARNING/ERROR yields nothing — even though the mirror has been silently failing for weeks.

After days of use, the Viking user/default/memories/ directory had only 2 entries (vs 30+ in built-in memory), confirming the mirror path never successfully completed.

Expected Behavior

on_memory_write should:

1. Handle all three actions: add, replace, remove
2. Use upsert semantics instead of create-only
3. Log failures at WARNING level so users can detect issues

Additional finding

There are also 3 stale openviking-server processes running (PIDs 30508, 401, 302) on this machine. This might contribute to resource is busy errors, but the root cause is the mirror code defects above.