Cross-session graph persistence (memory/ → graph merge)

[jeremiepas]

Problem

Conversations saved to memory/ are loaded on startup but NOT merged into the main graph.json. select_context queries only graph nodes, missing conversation nodes in structural queries.

Solution

On startup, merge conversations from memory/ directory into the loaded graph before building the index. After N mutations, write the enriched graph back to graph.json.

Startup flow:

1. Load graph.json
2. Load all conversations from memory/
3. Merge conversation nodes + edges into graph
4. Build index from enriched graph
5. Start MCP server with enriched state

Snapshot policy:

• Every 10 mutations: write graph.json atomically
• On SIGTERM/SIGINT: force snapshot
• On explicit snapshot MCP command: force snapshot

Acceptance Criteria

• On restart, previously saved conversations appear in select_context
• Periodic snapshot after 10 mutations
• Graceful shutdown writes final snapshot
• Atomic write (temp file + rename) to prevent corruption

Files to Modify

• src/Graphos/Infrastructure/Server/MCP.hs — merge conversations on startup
• src/Graphos/UseCase/Export.hs — add snapshot function
• src/Graphos/UseCase/Conversation.hs — add mergeConversationsToGraph

Effort: 1 day

Priority: Critical (depends on #5)

See: docs/proposals/memory-agent/02-gap-analysis.md (Gap 6)