feat: Memory compaction via LLM summarization (constructive, not destructive)

[alexeymorozua]

Problem

Current memory cleanup (if implemented at the agent level) is destructive: old or duplicate entries are deleted. This loses information — 5 related entries become 2 surviving entries rather than 1 dense entry containing the essence of all 5.

As memory modules approach their size limits, important historical context is dropped rather than compressed.

Proposed Solution

Adaptive Compaction: when a memory module exceeds a threshold, run an LLM summarization pass that compresses N related entries into 1 semantically-rich entry (reference: AgentRM paper, arxiv 2603.13110).

Before compaction (5 entries, ~40 lines):
  - Decision A made on 2026-01-10
  - Decision A revised on 2026-01-15  
  - Decision A applied to project X
  - Related decision B on 2026-02-01
  - Decision B supersedes part of A

After compaction (1 entry, ~5 lines):
  - [2026-02] Decisions A+B: <semantic summary preserving all key facts>

Trigger conditions

memory:
  compaction:
    enabled: true
    trigger_lines: 70        # compact when module exceeds this
    target_lines: 40         # compact down to this
    preserve_recency_days: 14  # never compact entries newer than N days

Difference from destructive cleanup

Approach	What happens to old entries
Destructive cleanup	Deleted if "stale" or "duplicate"
Adaptive compaction	LLM-summarized into denser form, semantics preserved

Compaction is complementary to cleanup — cleanup removes clearly obsolete facts, compaction preserves knowledge that's still relevant but verbose.

Implementation

Could be a built-in hook triggered after memory writes, or exposed as a framework utility callable from agent memory maintenance crons.