ayushmall
UserA personal memory layer for your AI tools. Intelligent authoring (gap detection, self-enriching graph, session-synthesis loop) + measurably good retrieval (94.9% blind Cov@10, sub-ms latency). Domain-agnostic framework — works for any context you want to track. MCP-native, Claude-Code-first, works with Cursor/Continue/Cline/OpenAI/Gemini.
Categories
Indexed Skills (18)
granola-series-recap
Cluster Granola meeting recordings into recurring series (e.g. "Weekly Jane <> Alex 1:1", "Platform weekly check-in"), create a granola-series surface entity per cluster, and synthesize cross-meeting decisions/project_facts from the series rather than treating each meeting as an island. Use when the user asks "what's been happening in my 1:1s with <person>?", "summarize the Platform weekly", "what decisions came out of recent <series> meetings?", or when batch-ingesting Granola as a recurring source.
memory-ask
Search the user's MemoryVault — their personal knowledge layer of memories (events, decisions, customer interactions, technical notes) linked to entities (people, companies, topics, projects). Use this BEFORE answering ANY question that references the user's work, past meetings, customers, colleagues, projects, or decisions they've made. The vault is the source of truth for the user's professional context. Returns top-K ranked memory snippets with titles, scores, entities, and content. Common triggers- "what did <person> say about <topic>", "what's our status on <customer>", "when did we decide <thing>", "tell me about <project>", "remind me what <person> wanted", "summarize last week with <customer>". Pass `k=10` for broad context, `k=3` for targeted answers.
memory-deep-dive
For a thin-retrieval queue item, fetch richer content from the source's native MCP, synthesize into a properly-shaped memory, save it. Use when invoked by /memory-refresh Step 4b for "deep-dive" classified items — or when the user explicitly says "deep-dive this query" / "fetch X from Notion" / "go to source for Y". Specialized per source (Notion, Slack, Linear, Gmail, GDrive, Granola, GitHub).
memory-entity-bootstrap
Enrich a low-info entity by pulling source data specifically about it. Triggers when ingest or refresh encounters an entity that's being newly created OR an existing entity with too few memories (< 5) that just got touched by new activity. Searches the user's enabled source MCPs (Slack history, Linear issues, Notion pages, Gmail threads, Granola meetings) for content about this entity, then synthesizes a richer canonical entity file or adds the missing context as new memories. Use when '/memory-refresh' wants to deep-dive a specific entity, or when 'memory-master-ingest' is about to create a new entity it should bootstrap properly first.
memory-graph-audit
Walk the user through a visual audit of their vault using Obsidian's graph view. Triggers on 'audit my graph', 'check structure', 'walk through vault visually', 'find graph issues', 'memory graph audit'. Some structural problems (orphan entities, duplicate-spelling entities, broken clusters, missing champions, dead wikilinks) are visible at a glance in Obsidian but invisible to the eval set. This skill gives the user a checklist of what to look for, captures their observations as actionable feedback memories, and accepts screenshots if they want to flag something visual. Pairs the model's structural checks (memory doctor) with human visual pattern-matching.
memory-ingest-code
Ingest code repositories into the MemoryVault as entities + PR memories. Use when the user wants their kit to understand their codebase — "ingest my repo", "pull PRs from <repo-name>", "add code context for <project>", "give me an engineer's memory", "set up code ingest for my monorepo" etc. Two safe modes: --metadata (README + structure only) and --prs (PR descriptions + paths only, never source contents). Source-content ingest is enterprise-only and disabled by default. Walks the user through products config setup for multi-product monorepos so PRs auto-link to the right product entities.
memory-master-ingest
The wide-net source scourer — iterates EVERY source the user has connected (defined in `<vault>/.mvkit/connected_sources.json`), invokes the right ingest, reports per-source status. THE most important Layer-1 agent. Called from /memory-refresh (Step 2), or directly when the user says "ingest everything" / "pull fresh data" / "what's new". The source list is data-driven — different users have different sources connected, so this skill reads the config at runtime instead of hardcoding which MCPs to call.
memory-refresh
The recurring entry point for an already-set-up MemoryVault. Triggers on 'refresh memory', 'pull fresh data', 'what is new', 'memory refresh', or just 'refresh' in any session where the kit is installed. Reads the vault state from .mvkit/, pulls deltas from every connected source, heals the graph, runs coverage detection, runs a soft eval, and reports what changed. Same skill works in any fresh chat because state lives in the vault, not the session. For first-time setup, invoke memory-setup instead.
memory-save
Persist a new memory to the user's MemoryVault. Use when the user explicitly says "save this", "remember that", "add to my vault", "note that down" — OR when you're summarizing a clearly memory-worthy outcome they just shared (a decision, a customer commitment, a new fact about a colleague). Writes a properly-formatted memory file with entity wikilinks. Don't use this for every utterance — be selective. If you're unsure, ASK the user "should I save this to your vault?" before calling. ALWAYS apply the preservation rules below — under-detailed memories are the single biggest quality failure mode of this system.
memory-setup
Conversational first-run setup. Triggers on "set up memoryvault", "install memoryvault", "initialize my vault", "I want to try memoryvault", or after a clone with "what's next?". Probes the user's environment for installed MCPs, asks which sources they want to ingest, gathers per-source config (Linear teams, repo names, Notion topics, Slack channels, etc.), writes `.mvkit/connected_sources.json`, generates the per-user scheduled tasks, walks through first ingest + first eval, and writes a `mem_BOOTSTRAP_*` audit memory. Every step gathers the info the kit needs so subsequent runs are automatic.
memory-stub-enricher
For a stub coverage-gap memory (tags include stub-enrich-me, enriched=false), read its auto-gathered Evidence section + any session context you have + write a grounded narrative via memory_update. Use when invoked by /memory-refresh Step 4b for "enrich-stub" classified items — or when the user notices a stub gap in retrieval results. Doesn't fetch from external sources — pure session-context-driven enrichment (memory-deep-dive is the escape hatch when no session context applies).
memory-use
The universal contract for ANY consuming agent using the MemoryVault MCP. Use this skill on every conversation that touches the user's professional context — meetings, customers, decisions, projects, code, anything from their daily work. Establishes when to refer to memory (search first, always), when to give feedback (mem_GAP_* enrichment + memory_update), when to write memory (memory_save discipline + playbooks), and when to escape to a native MCP for deep-dive (parent_surface tells you which source has richer info). Load on every client; the kit's MCP tool descriptions encode the rest.
memoryvault
Personal memory layer for non-engineers — runs entirely on Claude Cowork with Google Drive as the vault store. Reads/writes structured markdown memories about your meetings, decisions, customers, projects, and code. Use this whenever the user asks about their work history ("what did Maya say", "status of the SDK", "who's working on Domain"), or asks to save something ("remember that we decided X", "save this from the meeting"), or asks to pull recent activity ("refresh from yesterday's meetings"). Triggers on memory-related phrases — "memory", "vault", "remind me about", "what was the latest on", "save this", "remember this", "ingest", "audit my memory". Drive-backed, no install required. Six sub-flows internally: ask, save, refresh, audit, heal, ingest-code.
pylon-customer-history
Pull a customer's recent Pylon support history and synthesize it into structured memories tied to the customer's pylon-account surface entity. Use when the user asks "what's going wrong with <customer>?", "summarize <customer>'s Pylon", "what bugs has <customer> reported?", "what's the support backlog for <customer>?", or batch-ingesting Pylon as a recurring source. Each thread becomes a `type: feedback` or `type: project_fact` memory with `source_surface: "[[<customer-name> Pylon]]"`. Directly fills the G14 customer-triad gap (contact + meeting + commit) by capturing the support leg structurally.
slack-channel-digest
Pull recent activity from a Slack channel and write structured memories tied to the channel's surface entity. Use when the user asks "what's happening in
memory-audit
Get a one-shot health report on the user's MemoryVault — memory count, entity coverage, dead wikilinks, orphan entities. Use when the user asks "how's my vault?", "vault status", "is my memory healthy?", or before any deep work session where you want to verify the data layer is intact. Also useful BEFORE running memory-refresh to know if there's existing rot.
memory-heal
Auto-fix safe issues in the user's MemoryVault — backfill missing person aliases, resolve dead wikilinks by mapping to existing entities, mark orphan entity files as stubs. Use when memory-audit surfaces hygiene problems (dead_wikilinks > 0, entities_without_aliases growing) or once a month as preventive maintenance. Idempotent — running it twice is safe.
memory-ingest
Bulk-import a folder of markdown notes into the user's MemoryVault, OR add a single markdown file. Use when the user says "import my old notes from <path>", "bring in this file", or "convert this folder of notes to memories". Wraps each file with proper frontmatter, generates stable IDs, and lints. Different from memory-save (which is for single new memories from agent context).
Bio shown is the top-scored skill's repo description as a fallback — real GitHub bios land in a future update.