hermes - 💡(How to fix) Fix [Feature]: Memory usage frequency tracking for smarter eviction [1 participants]

Problem

The built-in MemoryStore (MEMORY.md / USER.md) has a hard character limit (memory_char_limit=2200, user_char_limit=1375). When the limit is reached, add() simply rejects the new entry and asks the agent to manually replace or remove.

There is no usage frequency tracking on memory entries, which means:

1. No data-driven eviction: When memory is full, the agent has to guess which entries are less important. There is no signal indicating which entries are actually referenced/used frequently vs. which are stale.

2. Equal weight for all entries: A frequently-referenced preference (e.g., "user prefers Chinese") occupies the same priority as a one-time observation from months ago. The agent cannot distinguish high-value from low-value entries.

3. Manual curation is error-prone: The agent must re-read all entries and make subjective judgments about what to remove. This is slow, inconsistent, and wastes tokens.

Proposed Solution

Add a lightweight use_count field to each memory entry:

On each replace or implicit reference:

• Increment use_count for the matched entry
• Store metadata in a sidecar file (e.g., MEMORY.meta.json) to avoid modifying the text format

On eviction (when add() would exceed limit):

• Sort entries by use_count (ascending)
• Suggest or auto-remove the lowest-frequency entry
• Optionally combine with recency (last modified timestamp)

Minimal implementation:

// MEMORY.meta.json
{
  "entries": {
    "用户太太：李杏元": {"use_count": 12, "last_used": "2026-04-26T10:00:00Z"},
    "软件安装规则：优先国内镜像": {"use_count": 3, "last_used": "2026-04-20T08:00:00Z"}
  }
}

Why This Matters

• Frequency = value signal: Entries that are referenced often are clearly more important. This is a much better eviction heuristic than position or age alone.
• Low implementation cost: A sidecar JSON file + increment on match. No format changes to MEMORY.md/USER.md.
• Complements #678: The Forgetting proposal covers decay/pruning. This adds the missing frequency dimension that makes decay smarter.

Related

• #678 — Memory Extraction + Forgetting (decay-based lifecycle)
• #11425 — Skills lifecycle management (similar frequency tracking for skills)
• #12987 — Ebbinghaus decay for mem0 plugin (external, not built-in memory)

Environment

• Hermes Agent latest (as of 2026-04-26)
• Built-in MemoryStore in tools/memory_tool.py