claude-memory/graph/solutions/cognitive-memory-v30-rich-edges-hybrid-embeddings-mcp-server-9ea720.md

id	type	title	tags	importance	confidence	created	updated	relations
9ea72015-0b85-4a42-9ae1-144866f8d86f	solution	Cognitive Memory v3.0: Rich Edges + Hybrid Embeddings + MCP Server	cognitive-memory mcp architecture upgrade	0.9	0.8	2026-02-19T20:10:42.128691+00:00	2026-02-28T06:03:25.212458+00:00	target type direction strength edge_id dc6b6c6f-1319-47cf-a691-dd40751ff121 BUILDS_ON incoming 0.85 78728fca-82c7-4c30-b677-aadc663fb40b target type direction strength edge_id 89f32ef3-2d87-4b8c-aa6e-ce5a7f60566c BUILDS_ON incoming 0.8 6100de8d-c1dc-4ec9-8981-5c65a396da0e

Major upgrade to cognitive-memory skill. Phase 1: Rich edges as first-class markdown files in graph/edges/ with bidirectional frontmatter refs (edge_id field). relate() returns edge_id string instead of bool. Cascade deletion on memory delete. CLI: edge-get, edge-search, edge-update, edge-delete. Phase 2: Hybrid embedding providers (Ollama local + OpenAI optional) with automatic fallback chain. _config.json stores provider settings (gitignored). Dimension mismatch safety triggers re-embedding on provider switch. Phase 3: MCP server (mcp_server.py) with 18 tools via JSON-RPC 2.0 stdio protocol. Registered in ~/.claude.json. Phase 4: Updated SKILL.md, SCHEMA.md, feature.json to v3.0.0. Index version bumped from 1 to 2 (adds edges section). All stdlib-only, no external dependencies.