hermes - 💡(How to fix) Fix Memory: auto-compact MEMORY.md when char limit reached [1 participants]

Problem

MemoryStore.add() in tools/memory_tool.py simply rejects writes when the char limit is hit (default 2,200):

# memory_tool.py:247-258
if new_total > limit:
    return {"success": False, "error": 
        f"Memory at {current:,}/{limit:,} chars. "
        f"Replace or remove existing entries first."}

There is no auto-compaction mechanism — no merging, summarizing, or FIFO eviction. Once full, MEMORY.md silently stops accepting new entries until the user manually cleans it up.

This is different from conversation context compression (compression.threshold), which works automatically but only touches the messages array, not memory files.

Proposal

When MemoryStore.add() detects the limit would be exceeded:

1. Option A (LLM compact): Fire a lightweight LLM call to merge/summarize existing entries, then retry the add. Like conversation compression, but for memory files.

2. Option B (FIFO evict): Drop the oldest entry, then retry. Simple but lossy.

3. Option C (compact-on-nudge): Let the background memory review (every 10 turns) also check usage percentage. If near full, add a compact instruction to the review prompt.

Option C is the least invasive — just add to _MEMORY_REVIEW_PROMPT something like:

If MEMORY.md is over 80% full, merge redundant or stale entries to free space.

Related

• Issue #5563 mentions the 2,200 char limit is critically small — this is the other half of the fix (not just bigger, but self-maintaining).