LongMemEval: improve multi-session and temporal conflict resolution

[jack-arturo]

Evidence

The canonical LongMemEval full run (benchmarks/results/longmemeval_full_gpt5mini_20260425_231308.json) scored:

• multi-session: 81.2% (108/133)
• temporal-reasoning: 88.0% (117/133)
• overall recall@5: 97.2% (486/500)

Problem

The system usually retrieves relevant evidence, but still misses questions requiring chronology, session ordering, or resolving changed facts across sessions. This points to ranking, metadata use, conflict resolution, or answer construction rather than broad retrieval failure.

Acceptance Criteria

• Analyze failed multi-session, temporal, and knowledge-update questions where recall_hit_at_5=true.
• Identify whether failures come from ranking, missing date/session metadata, outdated facts, graph expansion noise, answer construction, or another concrete category discovered during analysis.
• Add focused server tests for date-aware recall, latest-fact selection, and conflicting memories across sessions.
• Re-run the representative mini (--per-type 5) and full LongMemEval before claiming improvement.

Related Work

Related to existing recall/graph/metadata issues: #74, #100, #110, and #111.

[jack-arturo]

Cross-linking the focused current-state tracker: #169

The Memora/FAMA diagnostics in automem-evals produced a concrete instance of this issue: temporal/update questions can fail even when relevant memories are retrieved, because old, removed, or superseded facts are still treated as active recall context. That points to recall-time state resolution, not just higher recall@k.

The proposed slice for that issue is intentionally narrow: default /recall to active/current memories, suppress t_invalid/future t_valid/INVALIDATED_BY/EVOLVED_INTO stale nodes, and keep CONTRADICTS as evidence rather than suppression.