v0.5: L3 Reflective + L4 Corrective memory layers (self-learning + self-reflection)

[MakiDevelop]

Background

Per specs/08-memory-architecture.md, VirtualMe's full memory model has 4 layers. v0.4 implements L1 (episodic) + L2 (semantic). L3 (reflective) and L4 (corrective) are what make the system self-learning + self-reflective rather than "just an interview log."

What L3 enables

• Detecting contradictions between SOUL anchors and VOICE samples (e.g., spec says "direct" but actual samples to clients are diplomatic → register switching, not contradiction)
• Detecting cross-session drift in the extracted model
• Flagging coverage holes for next-session question selection
• Surfacing patterns across sessions for human review

Without L3, the system can extract and triangulate but cannot reason about its own modeling.

What L4 enables

• Capturing agent-mode rejections: "this isn't quite me — wrong tone / wrong vocabulary / factual error"
• Synthesizing lessons from rejections
• Auto-adjusting L2 (e.g., VOICE retrieval weighting shifts after consistent "tone too formal" feedback)

Without L4, the agent ships once and never improves from real-world use.

Implementation plan

Depends on issue #5 (MemoryBackend Protocol).

L3 — Reflective Memory

• Add reflections storage to backend Protocol
• Implement weekly reflection pass — scans past 7 days of L2 changes
• Reflection types: contradiction / drift / coverage_gap / pattern
• All L3 entries marked requires_human_review=True by default
• Operator UI: review and resolve reflections (CLI minimum)
• Pre-blind-test pass at Week 5 and Week 8

L4 — Corrective Memory

• Add feedback storage to backend Protocol
• CLI capture: `virtualme feedback --reject "draft" --reason "tone too formal"`
• Lesson extraction: LLM synthesizes "what to learn" from rejection
• L4 → L2 feedback loop: e.g., adjust VOICE retrieval weighting

Storage

memory-hall backend recommended (handles provenance + multi-agent attribution natively per spec/08). SQLite backend can accept writes but reflection passes are no-ops.

Acceptance criteria

• reflections table/namespace defined
• feedback table/namespace defined
• Weekly reflection pass runs (cron stub OK for v0.5)
• CLI feedback capture works
• At least one reflection type (contradiction detection) demonstrably catches a real case in integration test
• Documentation updated in specs/08-memory-architecture.md

Estimated effort

1 week. L3 alone is ~3 days; L4 capture + lesson extraction another ~3 days.

Why this matters

This is the difference between "VirtualMe extracts you once" and "VirtualMe learns about you over time." The 8-week interview methodology assumes the system gets better at modeling the interviewee as weeks accumulate — that requires L3. Post-ship improvement requires L4.