feat(palace): add MemoryProvider::recent for retention-ranked recall (mp-migration 4/8)

[quangdang46]

Summary

Add MemoryProvider::recent(limit, scope) — the retention-ranked recall used by jcode's recall --mode recent and the L1 wake-up layer.

async fn recent(
    &self,
    limit: usize,
    scope: Option<&SearchScope>,
) -> anyhow::Result<Vec<SearchHit>>
where Self: Sized;

Returns the top-N SearchHits scored by:

score = (access_count + 1) / (1 + days_since_last_accessed)

Drawers that have never been accessed fall back to "now" with access_count = 0.

Motivation

Required for the jcode → mempalace Mode C library migration (mp-migration 4/8). jcode has two call sites that need this:

1. MemoryTool::recall with mode = "recent" (no query) — tool/memory.rs:174-198
2. L1 wake-up layer in crates/jcode-base/src/memory.rs:856-880 (get_prompt_memories_scoped) — top-N most-relevant memories loaded into every prompt.

Both today compute a retention score by walking the full drawer list and sorting. With recent() on the trait, the adapter can delegate in one call.

This is PR 4/8 in the migration series. Independent of PRs 1/2/7 — recent only reads metadata["access_count"] and metadata["last_accessed"], doesn't require the typed fields or the mutation methods.

Change

crates/core/src/palace.rs — adds 1 default method to MemoryProvider (after get_drawers, before closing } of the trait). ~70 lines of logic + 30 lines of docs.

async fn recent(
    &self,
    limit: usize,
    scope: Option<&SearchScope>,
) -> anyhow::Result<Vec<SearchHit>>
where
    Self: Sized,
{
    // ... 70 lines of: pull → score → sort → trim → map to SearchHit
}

The scoring formula mirrors crate::retention::calculate_retention but operates on metadata fields rather than the typed Memory struct (which is the consolidate-pipeline's Memory, not the trait's Drawer). When the consolidation pipeline is wired to the trait, this default can be replaced with a direct call to kg.helpfulness_score(id).

Performance

• O(n) for embedvec (≤5 k drawers): one get_drawers call + in-memory sort.
• Pulls up to 4096 drawers (or limit.max(64) when a scope is set), then trims to limit after ranking.
• For larger palaces (usearch, lancedb) override with an indexed path that scores in the store.

Test plan

• cargo test -p mempalace-core --lib → 1127 passed, 0 failed (282s)
• cargo check -p mempalace-core clean
• cargo fmt --check -p mempalace-core clean
• CI green on ubuntu/macos/windows

Series status

#	Title	Status
6	docs(palace): document Embedder reuse on search_with_embedding	merged #16
7	feat(palace): add tags/trust/access_count/reinforcements/superseded_by to Drawer	open #17
8	feat(palace): add MemoryProvider::graph_stats_legacy (jcode shape)	open #18
1	feat(palace): add boost/decay/reinforce/supersede/set_metadata to MemoryProvider	open #19
2	feat(palace): add tag/untag/link/list_tags to MemoryProvider	open #20
4	feat(palace): add MemoryProvider::recent for retention-ranked recall	this PR
3	feat(palace): add MemoryScope::All and Wing/Room variants	next
5	feat(palace): add ActivityEvent sink on PalaceBuilder	next

🤖 Generated with Claude Code