MCP server, project registry, sectioned memories, two-tier tool design
Reduced MCP tool definition overhead from 2487 tokens (23 tools) to 1158 tokens (9 tools) while maintaining 100% CLI parity. 8 core tools are always advertised individually; 15 rare operations are accessed via a single memgraph_admin meta-tool that takes {command, args}. The admin tool's command enum tells the LLM what's available without paying for 15 full schemas on every API call.
Every CLI command (except daemons: serve, watch, mcp) now has a corresponding MCP tool. Agents can do everything via MCP that they could via the CLI — no shelling out needed. 18 new tools added (5 → 23 total): list, edit, delete, sessions, import, init, profile, status, config, attach, demo, bridge, setup, feedback, query, related, plans, graph_from_dir. Code split into 4 focused files (all under 500 LOC).
Added memgraph mcp command — a stdio JSON-RPC 2.0 server implementing the MCP protocol (2025-11-25). Agents like Claude Code auto-discover memgraph's capabilities without explicit CLAUDE.md rules. Register with: claude mcp add memgraph -s user -- ~/.local/bin/memgraph mcp. Initial 5 tools: memgraph_projects, memgraph_recall, memgraph_read, memgraph_save, memgraph_recommend. Eliminates the need for CLAUDE.md rules instructing agents to use memgraph.
Stable project identification system. Git remote URLs are now the primary scope key, with filesystem path fallback for local-only repos. A project registry (~/.memgraph/projects.json) maps human-readable names to memory paths. New memgraph attach <name> command registers, renames, and rebinds orphaned projects. Scope resolution order: --memory-dir (explicit) > --project <name> (registry) > remote-based > path-based. Backward compatible with old path-based scopes. Auto-imports existing scopes into the registry on first run.
Memories now support [slug] section markers that split multi-fact memories into addressable sections. memgraph recall returns a compact section index (slug + preview + line range) by default instead of dumping full content. Agents use memgraph read <id> <slug> to fetch only the section they need. Reduces token consumption by 60-90% for large memories. New memgraph read command for section-level retrieval. New memgraph list command for browse mode (no search query).
New memgraph projects command lists all project scopes across all repos with names, memory counts, and paths. Agents use this first when they don't know what projects exist. Works from any directory — no need to be in a specific repo.