--- name: memb-skill description: "BDB local-first long-term memory engine (memB). Query, remember, and adapt preferences, code architectures, and developer patterns across tasks." risk: low source: bdb date_added: "2026-07-11" --- # memB: Local Long-Term Memory Skill This skill allows agents to access and maintain a persistent, offline-first long-term memory bank using the **`memb-mcp`** server. It manages user preferences, project structures, and complex developer workarounds. --- ## 🔒 Memory Safety & Secret Filtration > [!IMPORTANT] > **Secret Ingestion Rule:** Under no circumstances should raw credentials, passwords, API keys (e.g. `GEMINI_API_KEY`, `POSTGRES_PRISMA_URL`), or raw environment configurations be written to memory. > Ensure all inputs are scrubbed of high-entropy strings before calling memory tools. --- ## 🛠️ Memory Categorization (Dynamic Flower-Like Graph Layout) `memB` implements a dynamic, project-agnostic hierarchical layout that structures memories into clusters automatically without hardcoded categories: 1. **"God Mode" / General Knowledge Hub (Center):** * Mapped using **`category="godmode"`** (with `project_id=None`). * Holds universal developer preferences, coding philosophies, global style sheets, and general core commands. 2. **Dynamic Project Leaves (Petals):** * Mapped using a specific **`project_id`** (e.g. `project_id="VisualSelect_By_BDB"` or `project_id="litha-gathering"`). * Isolates facts, custom configurations, and files routing patterns to the specific workspace project, preventing context pollution. * *System Integration:* The agent dynamically resolves the basename of the active workspace directory to scan and bind project memories automatically. --- ## 🔌 Using Memory Tools When working on tasks, query memory at the start of your turn to retrieve relevant developer context, and add critical decisions at the end. ### 1. Ingestion (`add_memory`) Use `add_memory` to commit a new fact, style preference, or design decision. * **Args:** `text` (string), `user_id` (string, default: "bdb_developer"), `category` (string, default: "godmode"), `project_id` (string, optional) * **Usage Guidelines:** * *General Preferences:* "Alice prefers to use absolute import paths globally." -> `add_memory({ text: "Prefers absolute import paths globally", category: "godmode" })` * *Project Learned Facts:* "In project VisualSelect, we must bypass the Firebase Auth login on localhost." -> `add_memory({ text: "Bypass Firebase Auth login on localhost", category: "project_node", project_id: "VisualSelect_By_BDB" })` ### 2. Retrieval (`search_memory`) Use `search_memory` at the beginning of a task to load context. * **Args:** `query` (string), `user_id` (string), `limit` (integer), `project_id` (string, optional) * **Behavior:** The search queries both global `godmode` memory and the active `project_id` memory in parallel, merging and ranking results by similarity. ### 3. Cleanup & Auditing (`list_memories` / `delete_memory`) * Use `list_memories` to audit active records. * Use `delete_memory` with a UUID to remove outdated or erroneous facts. ## 🚨 CRITICAL DIRECTIVE: AI-FIRST VAULT NAVIGATION 🚨 > [!CAUTION] > **DO NOT use arbitrary `find`, `ls -R`, or arbitrary file searches to discover project architecture!** > The `memB` ecosystem natively maintains a physical **AI-First Vault** at `~/.MemBDB/memB_Vault/`. > Treat `.openwiki` files as the highest-priority source of truth for architectural context. > 1. **Always read `~/.MemBDB/memB_Vault/God_Mode.md` FIRST** to understand the ecosystem topology. > 2. Navigate the tree via the `_Hub.md` files. > 3. Only search the actual workspace filesystem for raw code editing once you know the exact file path. --- ## 🧠 Agentic Ingestion (Intelligent Vault Building) You are responsible for intelligently categorized ingestion. Do NOT rely on static rules. When asked to ingest a project into memB: 1. **Analyze the Target:** Briefly scan the target directory to understand the project's purpose (e.g. 3D WebGL site, Python API, React Native app). 2. **Design Semantic Categories:** Dynamically invent highly precise categories tailored to the project (e.g., `3D_Engine`, `Routing_Logic`, `Database_Schemas`, `Styling_System`). 3. **Targeted Ingestion:** Execute the Python ingestion tool explicitly for specific files and categories, rather than doing a blind root scan. **Execution:** Use the `run_command` tool to execute `memb_ingest.py` on specific files or folders, explicitly passing the `--project` and `--category` flags. Because `memB` is supported across many AI platforms, the installation path depends on your current environment. The base path for MCPs is typically one of the following in the user's home directory (`~`): - `~/.gemini/mcps/` (Antigravity) - `~/.codex/mcps/` (Codex CLI / ChatGPT) - `~/.cursor/mcps/` (Cursor) - `~/.claude/mcps/` (Claude Desktop / Claude Code) - `~/.windsurf/mcps/` (Windsurf) - `~/.cline/mcps/` (Cline) - `~/.roo/mcps/` (Roo Code) - `~/.aider/mcps/` (Aider) **You must resolve the correct path first**, then run the ingestion command using the python binary from the `.venv` inside that `memb-mcp` directory. Example command structure (replace `` with the correct path discovered): ```bash ~//mcps/memb-mcp/.venv/bin/python ~//mcps/memb-mcp/memb_ingest.py path/to/specific_file.md --project "MyProject" --category "3D_Engine" ``` *(By injecting files one-by-one or in smart batches with precise categories, you build a flawless physical AI Vault that other agents can navigate intuitively).* --- ## 🚀 Micro-Targeted RAG (Task Execution) 1. **Task Start (Context Loading):** If `God_Mode.md` shows the project exists, read its Hubs. If you need highly specific snippets, use the `search_memory` MCP tool to query the vector DB. 2. **Execution:** Proceed with coding, applying the retrieved styles and preferences. 3. **Task End (Knowledge Capture):** If you resolved a complex setup bug or the user specified a new preference, run `add_memory` to persist it.