hermes - 💡(How to fix) Fix [Memory System] No importance-scoring mechanism — what should be remembered vs. forgotten? [1 participants]

 The Hermes memory system has a fundamental meta-cognitive flaw: **there is no mechanism to determine what information is worth preserving
 long-term vs. temporary context that should expire.**

 ---

 ## 1. No Importance/Relevance Scoring

 - **"Save everything" is the only strategy**: When in doubt, agents err on the side of saving. Over time, memory becomes bloated with low-value
 trivia while genuinely critical insights get buried.
 - **No "signal vs. noise" discrimination**: The system cannot distinguish between:
   - One-off observations (e.g., a specific error message)
   - Durable patterns (e.g., a user's working style, recurring failure modes)
   - Ephemeral context (current task state) vs. permanent knowledge (user preferences)
 - **No decay/importance recalculation**: Once saved, nothing ever gets promoted (important) or demoted (forgotten). Memory is write-only.

 **The paradox**: Users must manually tell the AI "remember this" — but users shouldn't have to manage memory. The AI should know what matters.

 ---

 ## 2. Fragmented Memory Surfaces (Architecture Problem)

 Four independent systems, each storing different types of information with no unified retrieval:

 | System | Content | Problem |
 |--------|---------|---------|
 | `memory` (memory.md) | Persistent notes about environment, conventions, tool quirks | Becomes bloated, 95%+ capacity |
 | `user` (user profile) | User preferences, communication style, personal context | Mixed with system-level facts |
 | `session/` | Full conversation transcripts | Never auto-summarized, grows forever |
 | `skills/` | Procedural knowledge (skills) | Separate from memory, no cross-reference |
 | gbrain/mempalace | External knowledge brains | Syncs separately, embedding failures cause silent degradation |

 **Result**: The AI has no single queryable memory — it must check multiple places with different interfaces, and the data is inconsistent across
 them.

 ---

 ## 3. Session Breaks = Context Loss

 - On reconnect, key context from before the break is lost unless it was explicitly saved to memory
 - What's actually preserved after a disconnect is unclear and unpredictable to the user
 - "I told you this last session" has no guarantee of working
 - `session_search` exists but is unreliable — searches miss relevant past sessions or return irrelevant ones

 ---

 ## 4. gbrain Sync Degrades Silently

 - Embedding failures (503 errors) cause gbrain to fall back to keyword-only search without alerting the user
 - The user perceives "search is working" when it's actually operating in a severely degraded mode
 - No health check, no warning, no self-healing

 ---

 ## 5. Skills and Memory Are Separate Silos

 - Skills store procedural knowledge (HOW to do things)
 - Memory stores facts/context (WHAT the user wants, WHY they want it)
 - These should inform each other but don't — no cross-pollination

 ---

 ## Root Cause

 This is a **meta-cognition problem**: the system has no internal model of "what is valuable to remember." It needs:

 1. **Importance scoring** — real-time assessment of whether a piece of information is worth long-term storage
 2. **Memory lifecycle management** — promote, demote, expire, consolidate
 3. **Consolidated retrieval** — one queryable interface, not 4 fragmented systems
 4. **Self-awareness of degradation** — know when gbrain/etc. is failing and tell the user

 ---

 ## Suggested Directions (Not Prescriptive)

 - Add a "memory pressure" metric that triggers consolidation when capacity is high
 - Tag each memory entry with a confidence/relevance score that decays over time
 - Build a unified memory API so the agent queries one place, not four
 - Add a `memory health` command showing which systems are degraded and how much data is "stale"
 - Auto-summarize old sessions into compressed memory entries

 ---

 *This issue captures systemic memory architecture problems identified through user feedback.*