Description

Implement the multi-view LTM perception Zoom-In pass from MemCoT (arXiv:2604.08216, #3564).

Extend ReasoningMemory::recall to run a Zoom-In localization step before the existing spreading activation. This step identifies where relevant evidence resides in long-context memory before expanding context.

Zoom-In returns the head edge plus its provenance/source message, reducing noise from chunk boundary splits. A companion Zoom-Out pass returns head edge plus 1-hop neighbors for causal context expansion.

Depends on SemanticStateAccumulator (#3574) being available as enriched query input — the accumulator's semantic state is the primary query context that Zoom-In refines.

Acceptance Criteria

• view: Head | ZoomIn | ZoomOut parameter wired through ReasoningMemory
• Zoom-In and Zoom-Out views covered by golden-file tests
• F1 measurement on internal LoCoMo fixture before/after (documented in playbook)
• memory.memcot.enabled = false disables Zoom-In completely (falls back to Head view)

References

• MemCoT paper: https://arxiv.org/abs/2604.08216
• Research synthesis: research(memory): MemCoT test-time memory scaling — multi-view long-context reasoning with episodic + semantic short-term memory #3564
• Depends on: feat(memory): add SemanticStateAccumulator to TurnContext for MemCoT semantic state tracking #3574 (SemanticStateAccumulator)
• APEX-MEM spec: /specs/004-memory/004-7-memory-apex-magma.md §14.1
• Implementation notes: .local/testing/playbooks/memory-research-2026-04.md