recall: respect invalidation and current-state semantics

[jack-arturo]

Problem

The Memora/FAMA diagnostic work in verygoodplugins/automem-evals exposed a recall-time current-state gap: AutoMem can retrieve relevant evidence while answer-level evaluation still fails because stale, removed, expired, or superseded memories remain active in /recall output.

This is not just a benchmark adapter concern. The graph already supports relation labels such as INVALIDATED_BY and EVOLVED_INTO, and memories already support temporal validity fields (t_valid, t_invalid), but /recall currently treats those as context rather than state semantics.

Evidence

Local diagnostic scope, not a public benchmark claim:

• A controlled FAMA-style smoke test seeded old/current/retracted facts and added INVALIDATED_BY, EVOLVED_INTO, and CONTRADICTS edges.
• /recall returned current facts but still leaked stale/retracted facts.
• A small Memora/FAMA answer-level slice showed that retrieval leakage diagnostics can undercount answer-level stale-fact failures.
• The local automem-evals workaround had to build a resolved-state ledger outside AutoMem to suppress superseded raw events before answer generation.

Related local artifact: automem-evals/data/results/memora-fama/20260522T140142Z-automem-relations-weekly-software_engineer/memora-fama-report.md.

Proposed behavior

Add recall-time current-state semantics so /recall defaults to active memories only:

• Suppress memories with archived=true.
• Suppress memories whose t_valid is in the future.
• Suppress memories whose t_invalid is in the past.
• Suppress memories with an outgoing INVALIDATED_BY or EVOLVED_INTO edge to an active replacement.
• Do not suppress on CONTRADICTS alone.
• If a suppressed memory has an active replacement, allow the replacement to be injected as a state_replacement result when it passes caller filters.
• Add aggregate diagnostics such as state_filter.suppressed_count and state_filter.replacement_count; include IDs/reasons only under an explicit debug flag.

Acceptance criteria

• /recall supports current_only defaulting to true and state_debug defaulting to false.
• current_only=false preserves legacy behavior for diagnostics/backcompat.
• Vector, keyword, tag-only, and relation-expanded recall paths apply the same current-state filter.
• CONTRADICTS does not suppress by itself.
• PATCH /memory/<id> preserves and syncs t_valid / t_invalid to Qdrant payloads.
• Tests cover expired, future-valid, invalidated, evolved, contradictory, and replacement-injection behavior.

Related work

• LongMemEval: improve preference memory update and recall semantics #158 and LongMemEval: improve multi-session and temporal conflict resolution #159 cover broader LongMemEval preference/temporal conflict behavior.
• feat: write-time dedup gate (ADD/UPDATE/SUPERSEDE/NOOP) #87 is complementary write-time dedup/supersede work, but recall-time suppression is still needed for existing stale graph state.
• End-to-end metadata audit: storage, consolidation, search, and recall #110/Return metadata fields in recall responses by default #111 cover metadata visibility and should remain separate.