Memory Consolidation and Pruning Schedule

PROJECT CONTEXT:
My agent's memory store grows monotonically: duplicates pile up, dead facts linger, and recall quality decays. Layered memory systems solve this with explicit lifecycle mechanics - consolidation runs, decay scores, and pruning with archives.

GOAL:
Implement a memory lifecycle: working/long-term layers, scheduled consolidation, decay-based pruning, and a safety archive.

REQUIREMENTS:
- Two layers: working memory (recent, volatile) and long-term (consolidated, stable)
- A scheduled consolidation run promoting worthy working entries to long-term
- Decay scoring from age, access, and confirmation signals
- Pruning that archives instead of deleting, with restore capability
- Health metrics: store size, duplicate rate, and recall hit quality over time

CONSTRAINTS:
- Consolidation merges and rewrites; it must never lose unique semantics
- Nothing is hard-deleted for 90 days - archives make pruning reversible

DEFINITION OF DONE:
- Working entries either promote or expire on schedule - the layer stays small
- Duplicate and superseded entries are measurably reduced by consolidation runs
- Pruned entries are restorable from the archive by ID or search
- Health metrics show a stable or improving store across simulated months

COMMON FAILURES TO AVOID:
- A single flat store aging into a landfill
- Consolidation summaries that quietly destroy unique facts
- Pruning by age alone, deleting rarely-used but critical knowledge
- No metrics, so decay is discovered only when the agent gets visibly dumber