memv-skills

Drop-in Claude Code skills + mirrored docs for building on mem[v].

10 skills auto-fire on relevant tasks. The mirrored docs sit on disk so agents can Read them lossless instead of fetching over the network. Refresh script keeps the mirror current.

For the design rationale and research citations behind every choice, see DESIGN.md. For the version history and bundled-doc snapshot dates, see CHANGELOG.md.

What you get

.claude-plugin/plugin.json     plugin manifest
skills/                        10 skills, each <100 lines
  memv-bootstrap/              first-time SDK install + auth setup
  memv-mcp-vs-sdk/             when to use MCP tools vs SDK code
  memv-add-memory/             write memories (add only — SDK has no update/delete)
  memv-search/                 semantic + graph-aware retrieval
  memv-spaces/                 tenancy, isolation, CRUD
  memv-files/                  file (PDF/doc/image/audio) ingestion
  memv-video-ingest/           video ingestion (mem[v]'s headline path)
  memv-graph/                  knowledge graph queries
  memv-mcp-setup/              install MCP clients / build your own
  memv-debug/                  error triage + sanity probes
docs/memv/                     mirrored docs.memv.ai (~420KB)
  README.md                    fat router — agents read this first
  llms-full.txt                single-blob, all docs concatenated
  llms.txt + _urls.txt         provenance
  core-concepts/, sdk/, mcp/, connectors/, support/, use-cases/, changelog/
scripts/sync-memv-docs.sh      refresh script (idempotent)
AGENTS.md.template             paste-in routing block for downstream projects

Install

Option A — Claude Code plugin (recommended)

In any Claude Code session:

/plugin marketplace add drPod/memv-skills
/plugin install memv-skills@memv-skills

Skills auto-discover. The 10 memv-* skills will fire on relevant tasks. The bundled docs sit in the plugin cache and are referenced from skill bodies.

Option B — Manual copy

git clone https://github.com/drPod/memv-skills.git
cd your-project
cp -R ../memv-skills/skills .claude/skills
cp -R ../memv-skills/docs/memv docs/memv
cp ../memv-skills/scripts/sync-memv-docs.sh scripts/sync-memv-docs.sh
chmod +x scripts/sync-memv-docs.sh

Then add the contents of AGENTS.md.template to your project's AGENTS.md (or CLAUDE.md).

Why this exists

mem[v] is small enough that agents could fetch docs ad-hoc, but every fetch costs a round-trip and an LLM-summarization pass that loses precision. A local mirror plus a hand-curated routing index (docs/memv/README.md) lets agents:

• Read lossless instead of WebFetch lossy
• Pick the right page by skimming a 6KB router instead of grepping
• Stay offline / version-pinned during a build session

The skills add a layer above that: procedural recipes that auto-fire when the agent hits a relevant task ("user wants to add a memory" → memv-add-memory skill loads → tells agent which docs to read + the space_id rule + the error-handling envelope).

Refreshing the docs mirror

mem[v] ships updates. Re-fetch + post-process:

./scripts/sync-memv-docs.sh

The script:

1. Pulls https://docs.memv.ai/llms.txt (index)
2. Pulls https://docs.memv.ai/llms-full.txt (all docs in one blob, ~160KB)
3. Pulls each individual .md page (35 files)
4. Strips the duplicated 3-line "Documentation Index" banner Mintlify auto-injects
5. Unescapes Mem\[v] → Mem[v]
6. Rewrites https://docs.memv.ai/<path> → docs/memv/<path>.md so agents Read locally instead of WebFetch-ing

Standing rules baked into skills

These hold across all 10 skills:

1. Spaces are isolation boundaries. Every write needs space_id. No global writes exist.
2. MCP says workspaceId. SDK says space_id. Same concept, different name. Translate when bridging.
3. Use SDK types directly. Don't invent wrappers.
4. No hand-rolled HTTP for mem[v]. SDK only.
5. Multimodal-first. Push raw text/files/video. Don't pre-process.
6. Read sdk/error-handling.md before any try/except.
7. MCP mcp__memv__* tools are dev-time inspection only — never from app code.
8. Never guess SDK shapes. If docs don't show it, flag — don't invent.

Token economics

• Naive (read all docs every task): ~165KB / ~40K tokens
• With memv-skills (skill + targeted Read): ~10–25KB / ~3–6K tokens
• ~6–10× cheaper per task, no info loss

Contributing

Issues + PRs welcome at https://github.com/drPod/memv-skills.

If mem[v] ships a docs change that breaks a skill, run ./scripts/sync-memv-docs.sh and submit a PR with the updated mirror + any skill text adjustments.