knowledge-worker

A personal knowledge graph that survives between AI conversations.
User-centered (not conversation-centered). Provenance-or-bust. Built on boring infrastructure.

Your AI is only as smart as what it remembers about you.

knowledge-worker is a local-first personal knowledge graph for carrying context across AI sessions. It turns notes into reviewable concepts, decisions, goals, and relationships, keeps source excerpts attached, and exports compact context you can paste into Claude, GPT, Ollama, or any other LLM workflow.

Your private graph stays on your machine, enabling you to preserve the thread of your own reasoning across AI sessions.

Why

AI conversations usually start from zero. You clarify a decision, name a constraint, sketch a goal, and then the next session forgets it. RAG can be heavy, full-note prompts are noisy, and most note apps do not plug cleanly into chat workflows.

Stop dumping context. Build memory. knowledge-worker turns chats, notes, decisions, and sources into a local provenance-backed knowledge graph, then uses graph analytics to show what matters, what connects, what is weak, and what context an AI should see.

knowledge-worker keeps the useful parts: cited claims, explicit relationships, human review, and a small context snapshot when you need continuity.

Quick Start

Requirements: Python 3.10+ on macOS, Linux, or Windows.

Install from PyPI

The core CLI has no runtime dependencies beyond the standard library. Optional extras pull in LLM backends and RDF export only when you need them:

pip install knowledge-worker                # core CLI, stdlib only (mykg / mygraph) 
pip install "knowledge-worker[rdf]"        # + Turtle/RDF export (rdflib)
pip install "knowledge-worker[anthropic]"  # + Claude-backed ingest
pip install "knowledge-worker[openai]"     # + OpenAI-backed ingest
pip install "knowledge-worker[ollama]"     # + local Ollama ingest
pip install "knowledge-worker[all]"        # all ingest backends + RDF exports

Verify the install (no clone needed — seed generates its own demo graph):

mykg --help
MYGRAPH_PATH=/tmp/knowledge-worker-demo.jsonld mykg seed
MYGRAPH_PATH=/tmp/knowledge-worker-demo.jsonld mykg summary

Using a virtual environment avoids Homebrew/system Python's externally-managed install errors:

python3 -m venv .venv
source .venv/bin/activate
pip install knowledge-worker
export MYGRAPH_PATH=/tmp/knowledge-worker-demo.jsonld 
mykg summary

What It Does

• Ingests markdown notes (generated from AI skill) into candidate graph nodes and edges.
• Generates pre-ingest deep-dive workspaces for sources that need synthesis before graph promotion.
• Requires provenance excerpts before claims become durable memory.
• Lets you review, accept, reject, or edit LLM proposals before merge.
• Searches by term, lists nodes by type, and finds paths between ideas.
• Exports an LLM-ready context snapshot for a fresh chat session.
• Audits memory shape with PageRank, betweenness, k-core, communities, weak claims, and provenance coverage.
• Generates an offline HTML graph viewer for exploration and demos.

How Memory Enters The Graph

Graph memory is promoted through a review lifecycle:

source note
  -> candidates.json
  -> validate
  -> review
  -> merge accepted items into MYGRAPH_PATH

Candidates are proposals, not memory. Validation checks schema, IDs, provenance excerpts, and edge endpoints. Review is the promotion gate. The active graph changes only after accepted candidates are merged.

The product contract is:

The model proposes. Artifacts expose reasoning. Provenance verifies. Human review promotes.

Claude or Codex App, No API Key

If you are already working with Claude, Codex, or ChatGPT in an app session, you do not need an API key. Ask the assistant to produce a *.candidates.json file that follows the schema in mygraph/extractor.py, then let the local CLI validate, review, and merge it. In Claude Code, the bundled /ingest-notes skill runs this flow for you:

python3 -m venv .venv
source .venv/bin/activate
python -m pip install -e .

mykg ingest path/to/your/notes.md --candidates-file path/to/your/notes.candidates.json

The app subscription helps you create the candidates file. The user (default) or AI gets to decide what needs to go in to the memory.

Design Principles

Provenance first. Every durable claim points back to a source document and literal excerpt.

Local first. The graph is a file on your machine. No cloud sync, accounts, or telemetry.

Review before merge. The LLM proposes. You decide. Deterministic validation runs before anything enters the graph.

Boring persistence. Compact JSON-LD until it becomes the limiting factor. The schema stays stable across storage backends.

Open-web storage. Local JSON-LD is the source of truth; JSONL records capture history, and Turtle/RDF exports let the graph participate in linked-data workflows without moving private memory into a hosted system. Additionally, JSON-LD, which has clear semantic markup gives AI engines like Google Gemini or Microsoft Copilot higher attribution confidence when extracting facts

For the shorter mykg command from a clone, install it editable inside a virtual environment:

python3 -m venv .venv
source .venv/bin/activate
python -m pip install -e .
MYGRAPH_PATH=examples/demo_graph.jsonld mykg query provenance

On Windows PowerShell:

py -3 -m venv .venv
.\.venv\Scripts\Activate.ps1
python -m pip install knowledge-worker

$env:MYGRAPH_PATH = "$env:TEMP\knowledge-worker-demo.jsonld"
mykg seed
mykg summary

From a clone, install editable instead:

python -m pip install -e .

$env:MYGRAPH_PATH = "examples\demo_graph.jsonld"
mykg query provenance
mykg audit --out "$env:TEMP\analytics.json" --html "$env:TEMP\memory_audit.html"

If PowerShell blocks activation scripts, run this for the current terminal session and activate again:

Set-ExecutionPolicy -Scope Process -ExecutionPolicy Bypass

Commands

Use Your Own Notes

You can ingest your notes with or without an API key.

Claude or Codex App, No API Key

If you are already working with Claude, Codex, or ChatGPT in an app session, you do not need an API key. Ask the assistant to produce a *.candidates.json file that follows the schema in mygraph/extractor.py, then let the local CLI validate, review, and merge it. In Claude Code, the bundled /ingest-notes skill runs this flow for you:

python3 -m venv .venv
source .venv/bin/activate
python -m pip install -e .

mykg ingest path/to/your/notes.md --candidates-file path/to/your/notes.candidates.json

The app subscription helps you create the candidates file. The repo still keeps graph validation and merge local.

Deep Dive Workflow

Use deep-dive when a source needs synthesis, challenge, or a reasoning workspace before it becomes graph memory:

mykg deep-dive notes.md --out-dir ~/private/deepdives/notes
mykg deep-dive inspect ~/private/deepdives/notes
mykg deep-dive add-to-graph ~/private/deepdives/notes

Generation creates a workspace with manifest.json, artifact-plan.json, Markdown artifacts, validation reports, an artifact-local graph summary, and canonical candidates. It does not mutate MYGRAPH_PATH.

add-to-graph reads the workspace manifest and delegates to the existing ingest validation/review/merge path. Keep using ingest directly when you already have a focused source note or hand-curated candidates file.

See Deep-Dive Interaction Model for the generate, inspect, challenge, approve, and add-to-graph semantics.

Automated API-Backed Ingest

If you want the CLI to call an LLM directly, use a provider API key or local Ollama.

For Anthropic API:

python3 -m venv .venv
source .venv/bin/activate
python -m pip install -e ".[anthropic]"
export ANTHROPIC_API_KEY=...

mykg ingest path/to/your/notes.md

The Claude backend also auto-detects Anthropic-compatible provider env:

• Anthropic API: ANTHROPIC_API_KEY or ANTHROPIC_AUTH_TOKEN
• Foundry: ANTHROPIC_FOUNDRY_API_KEY plus ANTHROPIC_FOUNDRY_RESOURCE or ANTHROPIC_FOUNDRY_BASE_URL
• Bedrock: AWS_BEARER_TOKEN_BEDROCK, or AWS credentials plus AWS_REGION/AWS_DEFAULT_REGION

For OpenAI API:

python3 -m venv .venv
source .venv/bin/activate
python -m pip install -e ".[openai]"
export OPENAI_API_KEY=...

mykg ingest path/to/your/notes.md --backend openai --model gpt-5.2

Graph Workflow

The public repo ships code, docs, and a fictional demo graph. Your real graph should live outside the repo or in the ignored default path, then be loaded explicitly:

MYGRAPH_PATH=~/my-private-graph/mygraph.jsonld mykg summary
MYGRAPH_PATH=~/my-private-graph/mygraph.jsonld mykg query "architecture"
MYGRAPH_PATH=~/my-private-graph/mygraph.jsonld mykg context

Your private mygraph.jsonld and/or legacy mygraph.json, generated private viewers, TTL/JSON-LD exports, eval logs, state logs, and local env files are ignored by default.

Storage Direction

In v0.8.0, knowledge-worker resolves the storage question this way: JSON-LD is the canonical local graph; JSONL is the append-only history layer for reviews, evals, analyzer runs, and future replay; Turtle/RDF export is the semantic-web bridge. Kuzu is a future optional local query backend, and graphify.net is a future publishing or interchange target.

See Storage Decision for the full decision record.

Memory Audit

mykg audit is a read-only layer over the graph. It ranks important concepts with PageRank, bridge ideas with betweenness, structural strength with k-core, communities with deterministic graph splitting, and weak claims from confidence and provenance gaps. It also includes directed idea-flow queues: idea_attractors for concepts that many edges point into, idea_generators for ideas that branch outward, and a weak_claim_queue that asks for human review actions instead of auto-promoting conclusions.

MYGRAPH_PATH=examples/demo_graph.jsonld mykg audit \
  --out /tmp/analytics.json \
  --html /tmp/memory_audit.html

The generated HTML puts ranked panels and legwork queues first, with the graph canvas second. This keeps the feature focused on memory governance instead of making the raw graph view the product.

Discovery Layer

Where audit ranks what the graph already says, mykg discover infers what it implies but does not yet say — and turns every inference into a reviewable proposal:

• Staleness radar: important nodes whose evidence trail has gone cold, scored by importance × days since the graph last touched them.
• Co-mention candidates: pairs that recur together across multiple sources but were never linked (CO_MENTIONED_WITH).
• Goal-alignment candidates: ideas and decisions structurally entangled with a goal they have no contribution path to (SERVES_CANDIDATE).
• Link prediction: Adamic-Adar over the semantic graph (RELATES_TO).
• Question debt: open questions ranked by age, centrality, and missing evidence; answered questions are detected via decision ABOUT edges.
• Corroboration: claims that hang on a single source (SINGLE_SOURCE).
• Bridge finder: cross-community connectors that remain after removing dominant hub "spines" that mask real bridges (BRIDGES).
• Tension detector: claims that are both supported and challenged, and goal contributions that inherit a challenge to the goal (TENSION_WITH).

MYGRAPH_PATH=examples/demo_graph.jsonld mykg discover \
  --out /tmp/discovery.json \
  --candidates /tmp/discovery.candidates.json

Discover never mutates the graph. Derived edges land in a candidates file — a promotion queue for human review. AI proposes, provenance verifies, the owner promotes. Committed sample output: examples/demo_discovery.json.

Contributing

See CONTRIBUTING.md. The core graph model is intentionally minimal; contributions that preserve that shape are preferred.

License

MIT