Nokori — Camada de Memória Comportamental para Agentes
KorenKrita/nokori
Transforma correções em regras que o agente lembra e aplica sozinho
Ver no GitHubNokori é uma camada de memória local que transforma correções em regras de comportamento duráveis para Claude Code e Cursor. Armazena tudo localmente em SQLite e recupera regras relevantes automaticamente durante sessões, bloqueando ações perigosas antes que o agente as execute.
Instalação
npx convez-ai KorenKrita/nokori1 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
Casos de uso
- Prevenir erros repetidos (force pushes, migrações esquecidas, comandos em produção)
- Manter preferências comportamentais consistentes entre projetos e sessões
- Bloquear automaticamente ações perigosas com base em histórico de correções
Veredito

Solução sofisticada e bem-arquitetada para um problema real: agentes esquecendo correções entre sessões. A abordagem local-first com SQLite é prudente, mas requer disciplina do usuário para alimentar o sistema com feedback estruturado. Ideal para desenvolvedores que trabalham com Claude Code/Cursor em workflows críticos.
README
A tradução para o português deste README ainda está na fila.
Nokori 残り
What experience leaves behind runs deeper than memory.
Nokori (残り) means what remains: the thing still standing in place after the noise dies down.
Every session ends, and every correction you made evaporates with it. In the next session the agent wakes a stranger again, the same stranger who force-pushes, forgets to run the migration, types a dangerous command straight at the production database.
Nokori refuses to let it forget. It settles every "don't do that" you ever said into recallable behavioral rules: when your words drift back toward that scene, the rule surfaces on its own inside the agent's context. New rules first live as candidates underwater, collecting evidence in the background. Only after the cold path and posthoc evidence trust them can the sharpest ones become Gate-eligible and block the first risky tool call before the agent touches your files.
Your data stays on your machine, in SQLite, the whole way through. Retrieval during a chat never touches a model. Only the post-session extract calls an LLM, and even then it is fed nothing but compressed session fragments. Want it fully offline? Point the endpoint at a local Ollama.
Who is it for
Before / After
| Without Nokori | With Nokori |
|---|---|
| The same correction is repeated every session | The correction becomes a durable behavioral rule |
| Risky tool calls rely on the agent remembering context | Trusted Gate rules can block before the tool runs |
| Preferences vanish with the chat window | Rules stay local and follow you across projects |
| Retrieval means waiting on a model | Hot-path recall is deterministic file I/O + scoring |
One minute overview
You correct Claude / Cursor
└─▶ Nokori carves a rule (what scene + what to do)
└─▶ Next time your words drift near that scene
└─▶ The rule auto-writes into the agent's context (reminder)
└─▶ If it later becomes trusted + gate_eligible:
block once before the first matching tool call (Gate)
During a chat Nokori only does retrieval and small file I/O, never making you wait on a model. The LLM is only called after the session closes, when it extracts new rules from the transcript at its own pace.
Quick install
Four commands. Local memory. No hosted database.
Prerequisites: Python >= 3.11, Claude Code or Cursor already installed
# Recommended: pipx with local semantic retrieval
brew install pipx && pipx ensurepath
pipx install "nokori[local-embed]"
# Register hooks
nokori install --all # or --cursor / default is Claude Code only
# Verify
nokori health
# Minimal install (BM25 only, no local model)
pipx install nokori
# Dedicated venv
python3 -m venv ~/.local/venvs/nokori
~/.local/venvs/nokori/bin/pip install "nokori[local-embed]"
echo 'export PATH="$HOME/.local/venvs/nokori/bin:$PATH"' >> ~/.zshrc
# From source
git clone https://github.com/KorenKrita/nokori.git && cd nokori
python3 -m venv .venv && source .venv/bin/activate
pip install -e ".[local-embed,dev]"
Full installation guide (Cursor config, updating, uninstalling) in Installation
Quick start
# 1. Add a candidate rule
nokori add \
--trigger "Force pushing to a shared branch" \
--action "Use --force-with-lease, or push to a new branch" \
--severity high_risk
# 2. Verify the shadow match
nokori test "I'll just git push --force this branch"
# 3. Run maintenance (let evidence move rules forward)
nokori maintain
# 4. Rule out of date? Dismiss it
nokori dismiss <short_id>
Just open Claude Code or Cursor and work as usual. When a rule matches, the agent sees the injected reminder before it replies. For trusted + gate_eligible rules, the first sensitive tool call is blocked once.
Core features
Documentation
| Guide | What it gives you |
|---|---|
| 🚀 Installation | pipx install, Cursor config, updates, uninstalling |
| 🧠 Architecture | flywheel mechanism, hook timing, injection vs Gate, Shadow Pool |
| ⚙️ Configuration | config.toml, environment variables, full reference |
| 🔎 Retrieval Engine | BM25, embeddings, RRF fusion, injection tiers |
| 🌱 Rule Lifecycle | state machine, promotion evidence, maintenance tasks |
| 🧊 Automatic Extraction | cold-path pipeline, merge strategy, async mode |
| 🛡️ Gate Mechanism | two-layer matching, configuration, prompt-hash safety |
| ⌨️ CLI Reference | all commands and options |
| 🖥️ Web UI | visual dashboard features and development |
Relationship with existing systems
| System | Relationship |
|---|---|
| CLAUDE.md | Complementary. Nokori doesn't touch your CLAUDE.md; it handles the dynamic "when X, do Y" |
| Claude Code auto-memory | No conflict. Memory leans factual, Nokori leans behavioral rules |
| Other memory plugins | Hooks can coexist, but avoid stacking many context-injection plugins |
Data storage
All data lives in one local directory, ~/.nokori/. There is no network sync. Rules store behavioral descriptions, not your source code. Only the cold-path extract calls an LLM; point the endpoint at a local Ollama for fully offline operation.
Development
python3 -m venv .venv && source .venv/bin/activate
pip install -e ".[local-embed,dev]"
python -m pytest tests/
Project constraints: hot-path hooks use only stdlib + urllib (no LLM calls between prompt and reply), all hooks wrapped in top-level try/except fail-open. Base install includes fastapi + uvicorn for the web dashboard.
License
MIT