convez.ai
MCP ServerProdutividadeAvançado
smixs

Autograph

smixs/autograph

Camada de memória tipada para agentes que escrevem em Obsidian

Ver no GitHub
BomPluginScore · de 0 a 100

Camada de memória tipada para agentes sempre ativos que escrevem em vaults Obsidian. Define uma schema única (schema.json) para taxonomia, decaimento de conhecimento, reparação de links e geração de MOCs, mantendo automaticamente a ordem em bases de conhecimento crescentes.

Instalação

$npx convez-ai smixs/autograph

1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.

Métricas

Estrelas
80
Forks
6
Acompanhando
0
Saúde
42%
Último commit
há 2 meses
Versão
—
Issues abertas
1
Contribuidores
3

Casos de uso

  • Manter ordem automática em vaults Obsidian alimentados por agentes IA
  • Gerenciar decaimento de conhecimento e remover informações obsoletas
  • Reparar links quebrados e detectar duplicatas em bases de conhecimento
  • Computar scores de saúde de notas e gerar Map of Contents (MOC)

Veredito

Ferramenta sofisticada para quem trabalha com agentes IA escribas em Obsidian em escala. Resolve o problema real de entropia em bases de conhecimento crescentes, mas exige Python 3.11+, integração com múltiplas plataformas de agentes (Claude Code, OpenClaw, Hermes, Codex) e expertise técnica para configurar a schema corretamente.

README

A tradução para o português deste README ainda está na fila.

🧠 autograph

A typed memory layer for always-on agents. One schema, one graph, any Obsidian vault.

Version Python Tests License

🇬🇧 English · 🇷🇺 Русский


An always-on agent writes notes into your Obsidian vault every day — voice transcripts, meetings, ideas, contacts. A month later you have 800 files, broken wiki-links, duplicates, and no one remembers whether status: ongoing is the same as status: active. autograph is the layer that keeps all of that in order automatically.

You describe the taxonomy once in schema.json: what card types exist (note, contact, project, CRM deal), which folder holds which domain, how fast each kind of knowledge decays, which statuses are allowed. From there the engine takes over — it places new files, repairs links, forgets cards you haven't touched for months, merges duplicates, computes a health score. One schema.json is read by Claude Code, OpenClaw, Hermes, Codex — any agent writing to your vault.

⚡ Install

/plugin marketplace add smixs/autograph
/plugin install autograph@autograph

Direct install without registering the marketplace:

/plugin install github:smixs/autograph

Update / uninstall:

/plugin marketplace update autograph
/plugin uninstall autograph
git clone https://github.com/smixs/autograph.git ~/dev/autograph
openclaw plugins install ~/dev/autograph
openclaw gateway restart

The skill lands in ~/.openclaw/skills/autograph/ and /autograph:research becomes available globally. For one workspace only, clone into <workspace>/.agents/skills/autograph/.

hermes skills install github:smixs/autograph/skills/autograph

The skill lands in ~/.hermes/skills/autograph/. Hermes doesn't support Claude Code slash-commands — commands/research.md is ignored, but the whole workflow lives in the skill, so just invoke the skill by name in chat.

Moving from OpenClaw to Hermes? hermes claw migrate imports autograph to ~/.hermes/skills/openclaw-imports/.

git clone https://github.com/smixs/autograph.git ~/dev/autograph
ln -s ~/dev/autograph/.claude-plugin ~/dev/autograph/.codex-plugin

Point Codex at ~/dev/autograph in your agents config.

🚀 Quickstart

# 1. Point at a vault — empty, chaotic, or already populated
/autograph:research /path/to/vault

# 2. Daily health check
uv run skills/autograph/scripts/graph.py health /path/to/vault

# 3. Decay pass: recompute relevance + tier for every card
uv run skills/autograph/scripts/engine.py decay /path/to/vault

# 4. Regenerate Map-of-Content indexes
uv run skills/autograph/scripts/moc.py generate /path/to/vault

The 10-phase bootstrap (discover → generate → swarm → enforce → cleanup → tag → dedup → link → MOC → verify) is in skills/autograph/references/bootstrap-workflow.md.

💡 Use cases

ScenarioCommandsWhy
Audit someone else's vaultdiscover.py → graph.py health → graph.py fix --applyUnderstand state before changing anything
Bootstrap empty or chaotic vault/autograph:research <vault>Interactive Q&A + explorer agent swarm → schema draft → approval
Create a card with guaranteed linksWorkflow 3 in SKILL.md: type → path → frontmatter → ## Related (hub + 2 siblings) → engine.py touchOrphan cards are dead knowledge — the skill won't close the task until links are in place
Import from CRM / external sourceengine.py init → enforce.py --apply → enrich.py tags --apply → enrich.py swarm-links --applyExports from HubSpot / Notion / OneNote / Apple Notes become indistinguishable from hand-written cards
Spaced repetition of forgotten knowledgeengine.py creative 5 <vault> + cronThe 5 oldest cards surface back in warm tier for review

⏰ Scheduling

Decay + health at night. Dedup + MOC on Sundays.

openclaw cron add --name "autograph-daily" \
  --cron "0 3 * * *" --tz "Europe/Amsterdam" \
  --session isolated --tools exec,read \
  --message "uv run ~/.openclaw/skills/autograph/scripts/engine.py decay /path/to/vault && uv run ~/.openclaw/skills/autograph/scripts/graph.py health /path/to/vault"

openclaw cron add --name "autograph-weekly" \
  --cron "0 4 * * 0" --tz "Europe/Amsterdam" \
  --session isolated --tools exec,read \
  --message "uv run ~/.openclaw/skills/autograph/scripts/dedup.py /path/to/vault --apply && uv run ~/.openclaw/skills/autograph/scripts/moc.py generate /path/to/vault"
hermes cron create "0 3 * * *" "autograph decay + health on /path/to/vault" --skill autograph
hermes cron create "0 4 * * 0" "autograph dedup + MOC on /path/to/vault" --skill autograph

Job output: ~/.hermes/cron/output/{job_id}/{ts}.md.

0 3 * * *  cd /path/to/vault && uv run ~/dev/autograph/skills/autograph/scripts/engine.py decay . >/tmp/autograph-decay.log 2>&1
5 3 * * *  cd /path/to/vault && uv run ~/dev/autograph/skills/autograph/scripts/graph.py health . >/tmp/autograph-health.log 2>&1
0 4 * * 0  cd /path/to/vault && uv run ~/dev/autograph/skills/autograph/scripts/moc.py generate . >/tmp/autograph-moc.log 2>&1

Targets: health ≥ 90, broken_links = 0, description coverage ≥ 80%, stale (>90d) < 20%.

🧬 How decay works

Ebbinghaus-style memory with three mechanisms, all configurable in schema.decay:

Each touch increments access_count. More retrievals slow forgetting:

strength = 1 + ln(access_count)
effective_rate = base_rate / strength
relevance = max(floor, 1.0 − effective_rate × days_since_access)

A card touched 5 times decays ~2.6× slower than one touched once.

TypeRateHalf-lifeRationale
contact0.005~100 daysPeople don't go stale quickly
crm0.008~62 daysDeals have medium lifecycle
learning0.010~50 daysKnowledge fades moderately
project0.012~42 daysProjects have deadlines
daily0.020~25 daysDaily notes lose relevance fast
default0.015~33 daysFallback for everything else

A touch promotes one tier at a time: archive → cold → warm → active. last_accessed is set to the midpoint of the interval — so without a re-touch the card naturally drifts back down.

📦 What's inside

autograph/
├── .claude-plugin/
│   ├── plugin.json         # Manifest (Claude Code, OpenClaw)
│   └── marketplace.json    # Marketplace entry
├── commands/research.md    # /autograph:research slash command
├── skills/autograph/
│   ├── SKILL.md            # Skill instructions for the model
│   ├── schema.example.json # Generic starting template
│   ├── references/         # Bootstrap workflow, card templates, schema docs
│   ├── scripts/            # 14 engine scripts (stdlib only)
│   ├── tests/              # 220 self-contained tests
│   └── evals/evals.json    # Skill-creator eval cases
└── LICENSE
ScriptPurpose
common.pyFM parser, walk, domain inference, decay formula
discover.pyPhase 1: scan vault, enum candidates
generate_schema.pyPhase 2A: discovery → draft schema
swarm_prepare.pyPhase 2B: bin-pack vault into agent batches
swarm_reduce.pyPhase 2B: consolidate + validate schema
research.py/research helper: gate + manifests + reduce
enforce.pyPhase 4: validate + auto-fix against schema
link_cleanup.pyPhase 5: remove phantom wikilinks
enrich.pyPhase 6/8: tags + catalog-based swarm-links via LLM
dedup.pyPhase 7: safe merge + .trash/
graph.pyHealth score, repair, backlinks, orphans
moc.pyMap-of-Content generation
engine.pyDecay, touch, creative recall, stats, init
daily.pyEntity extraction from daily/memory files

Requirements: Python 3.11+, uv, Obsidian-style vault (folder of .md with YAML frontmatter). Optional: OPENROUTER_API_KEY for enrich.py. No pip install needed — stdlib only.

Tests:

cd skills/autograph && uv run tests/test_autograph.py

🔗 Lineage

Grew out of smixs/agent-second-brain — a Telegram-first "second brain" bot that classified voice transcripts into an Obsidian vault with a 9pm daily report. The decay engine, vault-health scoring, and graph tools turned out to be the part every agent needed, not just that one bot. autograph extracts them into a shared memory layer for any runtime.


Made with care in Tashkent · MIT License · Issues

Ver o original no GitHub

Bom para

Relacionados