Research: reflection/consolidation loop + evaluation scenarios

[aurexav]

Parent

• Epic: Agent memory UX: MCP surface + skills + doc pointers (epic) #74

Problem

• Brain-like memory requires consolidation/reflection over time, not only storage + retrieval.
• We need a measurable loop that reduces context size without losing correctness.

Scope

• Define a consolidation policy surface (inputs, thresholds, caps, key strategy).
• Propose evaluation scenarios/datasets that show improvement (retrieval quality, prompt size, stability).
• Decide what runs agent-side vs server-side.

Acceptance criteria

• A short decision memo with recommended approach + tradeoffs.
• At least one evaluation scenario that can be implemented in elf-eval.

Notes

• Backlog draft: docs/plans/2026-02-23-agent-memory-mcp-skills-backlog.md.