Summary

Implements three research-backed memory improvements:

• research(memory): EverMemOS engram lifecycle — MemCell→MemScene consolidation for structured long-horizon recall (arXiv:2601.02163) #2332 EverMemOS MemScene consolidation — brain-inspired engram lifecycle. Background loop clusters semantically-related SQLite messages into MemScene records (migration 049). Scene members excluded from regular recall via MessageKind filter. LLM-generated label/profile per scene, label truncated to 100 chars.
• research: Focus autonomous context compression (agent-driven, 22.7% token reduction) (arXiv:2601.07190) #2218 Focus compress_context tool — agent-invokable native tool (not threshold-triggered). Knowledge block in ContextBuilder survives compaction. AtomicBool concurrency guard. CompressionStrategy::Autonomous variant.
• research(memory): A-MAC adaptive memory admission control — 5-factor value decomposition for principled fact filtering (arXiv:2603.04549) #2317 A-MAC admission control — write-time admission gate on memory_save path. 5-factor scoring with normalized weights. Fast path skips LLM when heuristic score exceeds threshold+margin. remember() return type changed to Result<Option<MessageId>> (breaking, pre-v1.0.0). Wired in AppBuilder, decisions logged via AuditLogger.

All new LLM-calling subsystems expose *_provider config fields. New [memory.admission] and [memory.tiers] scene fields added to config.

New config

[memory.admission]
enabled = true
threshold = 0.4
admission_provider = "fast"
[memory.admission.weights]
future_utility = 0.3
factual_confidence = 0.2
semantic_novelty = 0.25
temporal_recency = 0.15
content_type_prior = 0.1

[memory.tiers]
scene_enabled = true
scene_similarity_threshold = 0.80
scene_batch_size = 50
scene_sweep_interval_secs = 7200
scene_provider = "fast"

[memory.compression]
strategy = "autonomous"
compress_provider = "quality"

Breaking changes

• remember() returns Result<Option<MessageId>> (was Result<MessageId>) — all call sites updated
• remember_with_parts() returns Result<(Option<MessageId>, bool)> — all call sites updated

Test plan

• Verify cargo nextest run --workspace --features full --lib --bins passes (7074 tests)
• Enable [memory.admission] in testing.toml and verify A-MAC gate logs admission decisions
• Run a session with strategy = "autonomous" and call compress_context manually
• Verify scene consolidation loop starts with scene_enabled = true (log: scene consolidation: starting sweep)

Follow-up

• compress_provider runtime resolution for compress_context is deferred to a follow-up issue

Closes #2332
Closes #2218
Closes #2317