hermes - 💡(How to fix) Fix [Feature]: Tiered memory system (short-term, medium-term, long-term) for Hermes Agent

   - Attach a new tiered-memory module that can be enabled via config (e.g., memory.tiers.enabled: true).
   - ST-M: ephemeral in-memory store with automatic expiry (20s). “Remember everything” policy for the active session.
   - MT-M: in-memory cache with TTL 20 minutes. Includes lightweight compression/summary to fit budget.
   - LT-M: leverage existing memory tool API (memory.add/memory.replace, etc.) to persist validated items.
   - Data flow:
     - On each user input or assistant output, the system records to ST-M.
     - If an item meets promotion criteria (e.g., importance, frequency, recency), promote to MT-M.
     - If still valuable after MT-M retention, promote to LT-M via the memory tool.
     - Recall requests search ST-M, MT-M, then LT-M in order; results are ranked by tier and recency/importance.
   - Configurable knobs (example):
     - memory.tiers.enabled: true/false
     - memory.tiers.short.ttl_seconds: 20
     - memory.tiers.medium.ttl_seconds: 1200
     - memory.tiers.promotion_heuristic: [frequency, recency, explicit_user_tag]
   - API/interface (conceptual, to implement or wrap):
     - tiered_memory.remember(text, metadata, tier, ttl)
     - tiered_memory.recall(query, limit)
     - tiered_memory.promote(item_id, target_tier)
   - Implementation options:
     - Core integration: modify the agent’s memory/context pipeline to insert tiered storage and automatic promotion.
     - Extension approach: implement as a dedicated Hermes skill or memory manager wrapper that coordinates with the existing memory tool‑chain (lower risk, slower to deploy).
   - Observability:
     - Telemetry on hits/misses per tier, TTL expirations, average recall latency, memory footprint.

   - Lightweight scratchpad per session without MT-M/LT-M promotions (simpler but less powerful).
   - Rely solely on optimized prompt compression (less invasive but may not scale for long-running tasks).