Auto-load core memories into agent system prompt

[ramonlimaramos]

Problem

Memories saved with depth_layer: "core" (e.g. user preferences, hard rules, never-do-X feedback) only surface when an agent explicitly calls recall(). In long-running coding sessions, the agent often forgets to recall before each edit, so a saved preference like "minimize comments" gets ignored on the next task — even though it's marked critical.

Today this is mitigated only by a flat MEMORY.md file the harness auto-loads, but that bypasses Synapto entirely (duplication, drift, no semantic search benefit).

Proposal

Synapto auto-exports memories with depth_layer: "core" (and optionally a new pin: true flag) into the agent's persistent context. Two implementations possible:

1. Sentinel projection — Synapto maintains ~/.claude/.synapto-core.md (or equivalent) auto-regenerated whenever a core memory is added/updated. The harness loads this alongside MEMORY.md.
2. MCP-level injection — Add a list_core_memories() tool that the harness calls at session start and injects the result as a system reminder.

Why this matters

• Closes the "recall discipline" gap that requires agents to remember to remember.
• Lets feedback memories actually shape behavior without manual re-querying.
• Keeps Synapto as the single source of truth (no MEMORY.md drift).

Acceptance

• A memory saved with depth_layer: "core" survives session boundaries and is visible to the agent without explicit recall() calls.
• Removing or updating the memory propagates to the auto-loaded surface within one session.

Filed from a real session where a freshly-saved code-style preference was at risk of being forgotten on the next session.