Implement memory recency decay and access-based promotion

[neooriginal]

This issue is AI generated.

Problem

All stored memories are treated as equally relevant regardless of age, how recently they were accessed, or how often they are retrieved. A memory written 18 months ago and never referenced since surfaces with the same weight as one from last week that is retrieved every day. This means:

• Stale or superseded facts continue to pollute context.
• High-value recent memories are not prioritised over old noise.
• The memory store grows without any pressure to evict or downrank redundant entries.

Competitors use activation/decay models (similar to human working memory) where recent accesses raise a memory's retrieval score and passage of time without access lowers it.

Required change

• Add last_accessed_at and access_count columns to the memories table.
• Update these fields each time a memory entry is included in a retrieval result.
• Incorporate a recency + frequency score into the retrieval ranking alongside embedding similarity.
• Optionally: after a configurable dormancy window (e.g. 90 days without access), mark entries as low-priority and consider them for compaction or removal.

Acceptance criteria

• Memories accessed recently and frequently rank higher than identical-content older memories.
• Memory access counts and timestamps are recorded accurately.
• A memory that has not been accessed in the configured window is demoted or flagged for review.
• Tests verify that recency signal changes retrieval ordering.