Bug Description

When a session expires and triggers a proactive memory flush, if the gateway restarts during or after the flush, the same session gets flushed again because the _pre_flushed_sessions set is in-memory only and resets on restart.

Evidence

From gateway logs at 2026-03-25 04:00-04:05:

04:00:18 - Session 20260324_075052_df00a65e expired, flushing memories proactively
[Hermes Gateway Starting...]  ← Gateway restarted
04:04:51 - Same session expired again, flushing memories proactively (duplicate!)

The session was flushed twice, causing:

1. Duplicate LLM API calls via OpenRouter
2. Memory tool errors as the LLM tried to save already-saved memories
3. Wasted tokens and potential memory corruption

Root Cause

In gateway/session.py:

self._pre_flushed_sessions: set = set()  # session_ids already flushed by watcher

This set is initialized in __init__ and never persisted. When the gateway restarts, it resets to empty, so the expiry watcher in _session_expiry_watcher() sees the same expired session and flushes it again.

Proposed Fix

Persist _pre_flushed_sessions to state.db (the existing SQLite database used for other gateway state) so it survives restarts. The set should be:

1. Written to state.db after each successful flush
2. Loaded from state.db on gateway startup
3. Cleaned up when sessions are actually reset/deleted

Additional Issue

The flush prompt should also account for memory being near/full. Currently it just says "save memories" without giving guidance to consolidate when at capacity. When memory is >90% full, the LLM generates invalid tool calls trying to work around the limit.