Add MCP Memory Server for shared vector memory + knowledge context

[ttracx]

Implement a complete MCP-based shared memory system so the OpenClaw
Gateway and all nodes can persist and retrieve prior knowledge (chats,
channels, tasks, decisions, runbooks, artifacts) without repeating work.

Core components:

• services/mcp-memory: FastAPI MCP server with 7 tool endpoints
  (memory.write, memory.search, memory.get, memory.link,
  memory.summarize_scope, memory.attach_blob, memory.fetch_blob)
• PostgreSQL + pgvector for durable structured memory + vector index
• Redis for hot working-set cache (last 50 IDs, TTL 6h) and search
  result caching (TTL 10min)
• bellie-blobnlie blob store integration for large attachments
• Hybrid retrieval: vector similarity (0.75) + trigram keyword (0.25)
  with metadata filters and strict tenant+scope isolation
• Content-hash based idempotency and dedupe
• Gateway middleware hooks for automatic memory writes on message
  received and task/tool completion
• Node preflight context bootstrap for injecting prior knowledge
• Scheduled jobs for scope summarization, memory promotion, and
  chat_turn pruning
• SQL migration with pgvector/pg_trgm extensions, 6 tables, and
  optimized indexes (IVFFlat, GIN trigram, GIN tags)

Integration updates:

• Root docker-compose.yml: mcp-memory as opt-in profile (--profile memory)
• Makefile: memory-up, memory-down, memory-logs, memory-health, memory-test
• GitHub Actions: build+push mcp-memory image, run mcp-memory tests in CI
• devcontainer.json: forward MCP ports, add docker-in-docker feature

https://claude.ai/code/session_015uJfi7Exmb3Ev4xwiAZY9D