--- name: vektor-memory description: > Full VEKTOR persistent memory skill. Auto-recalls context at session start, stores decisions and facts during work, searches memory on demand. Use whenever vektor MCP tools are connected. Triggers: remember, recall, store, memory, what did we decide, context, briefing, forget, search memory, vektor, percept. --- # VEKTOR Memory Skill — v1.5.8 > Requires VEKTOR MCP tools connected via Claude Desktop, Cursor, Cline, Roo Code, > Windsurf, or any MCP-compatible host. Licence key must be activated. > Download and setup: https://vektormemory.com/downloads --- ## Session Start — Always Run First At the start of every session, call vektor_briefing before responding. This loads your personalised context digest: recent decisions, active projects, preferences, and unfinished threads so Claude starts informed, not blank. --- ## When to Store Memories Call vektor_store whenever something is worth remembering across sessions: - User states a preference: tone, format, tool, workflow - Decision is made: architecture, design, approach selected - Important fact surfaces: config path, API key location, domain fact - Task completed: what was done, files changed, outcome - Error resolved: what broke, root cause, fix applied - Project context: goals, stack, team, constraints Example: vektor_store({ content: "User: TypeScript strict mode, interface not type, no semicolons, 2-space indent.", importance: 4, tags: ["preference", "typescript", "code-style"] }) Store after significant exchanges, not every message. High signal only. --- ## Recalling Memory Semantic recall (best for open context): vektor_recall({ query: "what did we decide about the database schema" }) Ranked fusion recall — BM25 + semantic, most accurate: vektor_recall_rrf({ query: "authentication approach" }) Search with filters: vektor_search({ query: "deployment", tags: ["devops"], limit: 5 }) Use recall when: user asks what was decided/discussed, starting a topic covered before, or needing project-specific facts before giving advice. --- ## Memory During Long Tasks Checkpoint frequently for multi-step tasks: 1. Start: vektor_briefing — load context 2. During: vektor_store after each meaningful step 3. End: store summary of what was done and next steps Example end-of-task store: vektor_store({ content: "Completed: JWT auth refactor. Next: update middleware. File: src/middleware/auth.ts", importance: 3, tags: ["task", "auth", "next-steps"] }) --- ## Ingesting Documents and Files vektor_ingest({ content: "", source: "project-requirements-v2.pdf", tags: ["requirements", "project-alpha"] }) Use for: PRDs, meeting notes, codebases, research papers, spec docs. Also use vektor_ingest at session end to ingest the full conversation history. --- ## Memory Delta and Briefing Recent changes (last N hours): vektor_delta({ hours: 24 }) Structured daily briefing: vektor_briefing() Useful for: morning context load, catching up after a break, reviewing recent stores. --- ## Memory Graph vektor_graph({}) Opens an interactive graph of memory clusters. Useful for spotting gaps or visualising how topics connect across sessions. --- ## Percept — Ambient Memory (v1.5.8) Percept surfaces memory contextually as you work without being asked. Morning handover on REPL startup — surfaces yesterday's active thread: Automatic — runs on vektor chat startup if Percept Chat is active. Drop files into inbox for background processing: vektor percept --once (process inbox, then exit) vektor percept (daemon mode, watches continuously) Supported drop formats: .txt .md .json .transcript .log Processed files move to ~/.vektor/inbox/done/ Depth levels (surfaced contextually, never pushed): Whisper — single concept word, barely visible Bloom — short phrase, one connection Anchor — synthesised insight, held until topic shifts --- ## Best Practices - Be specific when storing. "User wants clean code" is weak. "User follows Airbnb ESLint config, no semicolons, 2-space indent" is strong. - Tag everything. Use: project name, topic, type (preference, decision, fact, error, task). - Recall before advising. On any technical topic, run vektor_recall first. - Trust the briefing. vektor_briefing replaces "catch me up" prompts. - Use vektor_recall_rrf for precision. Dual-channel BM25 + semantic beats recall alone. - Memory compounds. VEKTOR re-ranks memories by usage patterns over time. - Importance scale: 1=trivial, 3=default, 4=important, 5=critical. --- ## Tool Quick Reference | Tool | Purpose | |---------------------|------------------------------------------------------| | vektor_briefing | Load personalised session digest | | vektor_store | Save a memory (fact, decision, preference, task) | | vektor_recall | Semantic similarity search | | vektor_recall_rrf | Ranked fusion recall — BM25 + semantic (most accurate) | | vektor_search | Search with tag, date, importance filters | | vektor_ingest | Load a document or large text block into memory | | vektor_delta | See what changed in memory recently | | vektor_graph | Visual memory map | | vektor_status | Memory DB health check | | vektor_text | Generate text using memory as context | | vektor_image | Generate images (MiniMax, OpenAI) | | vektor_speech | Text-to-speech | | vektor_vision | Analyse images with vision models | | vektor_providers | List configured media providers | | vektor_context | Assemble formatted context block for prompt injection| | vektor_profile | Synthesise user/agent profile from stored memories | | vektor_related | Traverse memory graph edges for a specific memory | | vektor_latency | Benchmark recall latency (p50/p95/p99) | | tokens_saved | Check how much context VEKTOR compressed this session| --- ## Installation 1. Download VEKTOR CLI: https://vektormemory.com/downloads 2. Run: vektor setup (connects Claude Desktop, Cursor, Cline, Roo Code, Windsurf) 3. Drop this file into .claude/skills/ in your project or global skills directory 4. Start a new session — VEKTOR auto-loads Same licence key works across all VEKTOR products. Docs: https://vektormemory.com/docs Percept: https://vektormemory.com/docs/percept Changelog: https://vektormemory.com/docs/changelog