pi-noodle

Long-term memory for Pi.

This repo is building a small, opinionated memory system that tries to be:

• useful — retrieve facts that actually help future turns
• safe — avoid saving temporary or sensitive content
• automatic — capture durable signals from normal conversation
• inspectable — review what was saved, pending, or discarded

Under the hood it uses libSQL for storage and vector similarity search for retrieval.

Quick start

# Install as a Pi extension
pi install @brianmichel/pi-noodle

# In Pi, configure interactively
/noodle settings

The setup screen shows the full config on one page:

1. Database mode — local file or Turso Cloud
2. Embedding provider — OpenAI, LM Studio, Ollama, or custom
3. Relevant fields update in place as you switch modes/providers
4. Required fields are validated before save

Settings are saved to ~/.pi/noodle/config.json — memories travel with you across all projects.

/noodle command

/noodle                  Show current config (paths, endpoint, masked API key)
/noodle remember <text>  Save a memory directly
/noodle forget <query>   Find and delete a memory
/noodle edit <query>     Find and update a memory
/noodle review           Review recent auto-saved memories
/noodle settings         Interactive single-screen configuration editor with validation
/noodle setup            Alias for /noodle settings
/noodle init             Create a default config file for manual editing
/noodle web              Start the Memory Explorer (auto-stops when all tabs close)
/noodle web stop         Stop the explorer immediately
/noodle web dev          Dev mode — hot reload on save, use web stop when done
/noodle web 8080         Start on a custom port

For UI development outside Pi, run npm run web:dev from the repo — it connects to your configured database and reloads the browser whenever you edit src/web/index.html.

Memory Explorer Web UI

Launch a dark-themed web interface to browse, search, and visualize your memories:

• Live stats — total memories, categories, scopes
• Category filter — dropdown of all stored categories
• Text search — substring matching on memory text
• Dark mode — GitHub-inspired color scheme

Run /noodle web in Pi to open the explorer in your browser. The server runs in a background process and shuts down automatically ~2 seconds after you close all tabs. Use /noodle web stop to kill it manually.

Config file

~/.pi/noodle/config.json:

{
  "db": {
    "mode": "local",
    "path": "/Users/you/.pi/noodle/memories.db"
  },
  "embedding": {
    "provider": "openai",
    "apiKey": "sk-...",
    "baseUrl": "https://api.openai.com/v1",
    "model": "text-embedding-3-small"
  }
}

Cloud mode (Turso)

{
  "db": {
    "mode": "cloud",
    "url": "libsql://my-db-org.turso.io",
    "authToken": "eyJ..."
  },
  "embedding": {
    "provider": "openai",
    "apiKey": "sk-...",
    "baseUrl": "https://api.openai.com/v1",
    "model": "text-embedding-3-small"
  },
  "extractor": {
    "mode": "balanced",
    "model": "deepseek/deepseek-v4-flash:free",
    "triggerEvery": 10
  }
}

Memory modes

When memory mode is not off, capture uses a unified capture pipeline with policy-gated persistence:

• conservative
  • fewer extraction runs
  • higher bar for auto-save
  • softer inferences are usually discarded until reinforced
• balanced (recommended default)
  • durable facts can auto-save
  • medium-confidence preferences go to /noodle review
• proactive
  • more frequent extraction
  • more candidate discovery
  • pending review queue grows faster, but saved-memory safety rules stay the same

The extractor model is configurable too, so you can tune quality/speed/cost separately from behavior mode.

Environment variable overrides

Env vars take priority over the config file:

Variable	Overrides
NOODLE_CONFIG_PATH	Config file location
NOODLE_DB_PATH	Local DB path
NOODLE_DB_URL	Cloud DB URL
NOODLE_DB_TOKEN	Cloud DB auth token
OPENAI_API_KEY	Embedding API key
EMBEDDING_BASE_URL	Embedding endpoint URL
EMBEDDING_MODEL	Embedding model name
NOODLE_EXTRACTOR_MODE	Memory mode: off / conservative / balanced / proactive
NOODLE_EXTRACTOR_MODEL	Extractor model ID
NOODLE_EXTRACTOR_TRIGGER_EVERY	Automatic extraction cadence in user turns
NOODLE_EXTRACTOR_DEBUG	Show the extractor debug widget: true / false

Architecture

Pi lifecycle events
  └─► MemoryService.capture(event)
        ├─► heuristic capture
        ├─► optional LLM extraction
        ├─► candidate promotion policy
        └─► conversation capture / consolidation when needed
                    │
                    ▼
              MemoryBackend
                    │
               TursoBackend
               ├─► libSQL (local or cloud)
               └─► Embedder

Capture pipeline

input / compact / switch / shutdown
  │
  └─► MemoryService.capture(event)
        │
        ├─► heuristic prefilter
        │      ├─ blocks sensitive / temporary content
        │      └─ catches explicit memory asks
        │
        ├─► optional LLM extraction
        │      └─ turns conversation context into memory candidates
        │
        ├─► local promotion policy
        │      ├─ save     → durable memory DB
        │      ├─ pending  → /noodle review only
        │      └─ discard  → dropped
        │
        └─► retrieval injects only saved memories

Why it is shaped this way

• Pi only tells the memory system what event happened
• MemoryService decides which capture stages should run
• heuristic and LLM candidates share the same promotion path
• pending memories stay out of retrieval until reviewed or reinforced

What gets stored

Every memory is a row in SQLite with text, embedding (F32_BLOB), category, categories, scope (userId/assistantId/sessionId), and arbitrary metadata.

Search

Vector similarity via vector_distance_cos() in libSQL, ranked by cosine distance, post-filtered by category and threshold.

Policy and review

The system intentionally separates:

• detection — heuristics and LLM extraction find memory candidates
• promotion — local policy decides save / pending / discard
• retrieval — only saved memories are injected into prompts

That keeps the system proactive without letting low-confidence guesses pollute retrieval. Pending candidates stay visible in /noodle review but are not injected until promoted.

File layout

src/
├── config.ts              # Config resolution (~/.pi/noodle/config.json + env vars)
├── config-screen.ts       # Flat single-screen config editor for /noodle settings
├── constants.ts           # DEFAULT_AGENT_ID
├── types.ts               # NoodleConfig, JsonObject, NotificationTarget, etc.
├── utils.ts               # maskSecret, describeError, formatJson, extractTextContent
├── commands.ts            # /noodle command + interactive setup entrypoint
├── extension.ts           # Pi extension lifecycle hooks
├── tools.ts               # memory_add / search / list / get / update / delete
├── session.ts             # Session message collection
├── queue.ts               # Sequential async write queue
├── notifications.ts       # UI notification helpers
└── memory/
    ├── backend.ts         # MemoryBackend interface
    ├── types.ts           # MemoryRecord, MemoryScope, etc.
    ├── turso-backend.ts   # TursoBackend (libSQL + vector search)
    ├── embedder.ts        # Embedder type
    ├── embedders/         # openai.ts, lm-studio.ts
    ├── service.ts         # MemoryService (event-driven capture pipeline + promotion)
    ├── policy.ts          # Heuristics (classification, repetition, retrieval)
    └── runtime.ts         # Wiring (config + TursoBackend + MemoryService)