Learning path

Separate a few objects first, then finish one path that matches your goal.

You do not need to read the whole site first. Separate a few objects, then pick the path that matches the work in front of you.

Do not walk several paths at once. A local API loop, a Python framework, and an agent plugin answer different questions. Finish one, then decide whether you need the next.

1. Separate the objects first

   Start with the mental model overview, then read Scope and fail-open. At minimum, keep Source, Memory, PreparedContext, and Handoff apart.

2. Start a local Server

   Follow Install and run the Server with the minimum setup. Writing Memory and using full-text search do not need an inference provider.

3. Choose one path

   Pick one goal below. Finish that loop before comparing other frameworks or agents.

4. Use official docs for details

   For install options, interface details, and deployment, use the latest published official docs. PowerContext 101 helps you choose a path and verify the result.

The documented installation targets are macOS and Linux with Python 3.11 or later. This tutorial does not provide an unverified Windows workaround.

Choose by your goal

I am new to PowerContext

Read in this order:

1. Mental model overview
2. Scope
3. fail-open
4. Install and run the Server
5. API tour

How do you know it worked? Follow steps 4 and 5 above. Start the local Server on Install and run the Server, then open the API tour and run the script to write one Memory.

The terminal should show three things: search_memory hits that Memory, prepare_context returns non-empty content, and the result includes a Citation. If it does not match, start with Troubleshooting.

Skip framework pages, agent-host pages, and the later Source / Handoff tutorials for now.

If you want a little more (optional)

If the objects still blur, read Source, Memory, Handoff, and Experience and Skill. After the first loop, continue only if you need Source and Memory, the Handoff loop, or Experience / Skill Review. To connect an MCP client to the same data, read Connect with MCP. To check active and inactive counts for the same Scope, read Verify Scope statistics in the Dashboard. Read Memory text through the Client, HTTP, or MCP.

I write Python agents

Separate the objects, then install and run the Server, finish the API tour, and open Framework integration.

How do you know it worked? First run the API tour script until you see a write, a search hit, and a cited prepare_context result. Then open the framework integration overview and connect that same Memory in the LangGraph or LangChain page. A new session should still find the conclusion you just wrote.

For a first integration, choose LangGraph or LangChain. Do not start with Server inference or a Preview adapter. The Pydantic AI Agent adapter is Preview and has no working standalone install path, so it is not a first integration.

If you only need automatic recall, you can leave Handoff and Skill for later.

If you want a little more (optional)

After LangGraph or LangChain works, read framework tricks. Add Source and Memory, the Handoff loop, and Experience / Skill Review only when you need extraction, Handoff, or Skill. Read Server inference and the Preview adapter later, not first.

I use Hermes / OpenClaw

Separate the objects, then install and run the Server. Open the common agents overview, then Hermes or OpenClaw.

How do you know it worked? Open the Hermes or OpenClaw page and finish its install plus powercontext doctor <host>. A passing doctor only means the plugin is installed. Open a new session and ask for the Memory you stored explicitly. If the session can find it, the path works.

After powercontext setup hermes, run hermes memory setup and select PowerContext. For OpenClaw, confirm that the plugin owns the Memory slot. WorkBuddy uses its own setup workbuddy path. The portable Agent Plugin has no setup command; load its package directory and verify explicit MCP operations. Installation, credentials, and durable writes still require your authorization.

If you want a little more (optional)

After recall works, read Handoff, the Handoff loop, and Experience / Skill Review. Agent self-extension covers managed Skill export only. Hermes and OpenClaw cannot export; switch to Codex or Claude Code for that step.

I use a coding agent

Separate the objects, then install and run the Server. Open the coding agents overview, then the host page you need.

Choose Codex, Claude Code, DSH, OpenCode, or Pi. They share the same Server. Their lifecycles, injection points, and permission checks differ.

How do you know it worked? Open the coding agents overview, pick a host, and follow that page’s install and checks. After you stop the Server, ordinary coding should still continue. If an explicit Memory write or Handoff commit fails, the page or terminal must show the error. Each host page rewrites these steps with its own tools.

If you want a little more (optional)

After the host page works, read Hooks and plugins and Extend hooks. Use Source, Memory, Handoff, and Experience and Skill for object detail. The extra runnable pages are Source and Memory, the Handoff loop, and Experience / Skill Review.

I am building an integration

Finish Hooks and plugins, compare at least two existing hosts, then read Extend hooks.

The deliverable is not a snippet that can call prepare_context. Write a verifiable integration note. Scope, temporary injection, Source capture, permissions, diagnostics, installation, and real-host end-to-end checks all need explicit behavior.

You can leave the Handoff and Skill tutorials until recall and capture are stable.

If you want a little more (optional)

After the main chain is stable, add Source and Memory, the Handoff loop, Experience / Skill Review, and framework tricks. The object pages help you check behavior. They are not a gate for starting: Source, Memory, Handoff, and Experience and Skill.

Check it with the same question

Whatever path you pick, use this to see if it actually works. On the first-contact path, do it in the API tour. After you connect an agent, redo it on that host page with its own tools.

You just wrote a conclusion you will need later. Can the next request find it?

1. Write it explicitly in a fixed Scope.
2. Ask again in a new request or session.
3. Check that the recalled content has a Citation and enters only this model request.
4. Stop the Server and ask again. Automatic recall may fail. The original task should continue.
5. Try one explicit write. The failure must be visible. It cannot pretend the save succeeded.

Do not treat this list as a checkbox by itself. If the symptom does not match, start with Troubleshooting.

If PowerContext fails while automatically fetching context, keep working. If you explicitly write a Memory, commit a Handoff, approve a proposal, or export a Skill, you must see the failure. Read fail-open for the full rule.

PowerContext 101 does not keep a separate Bub path. Its notes live mainly in the integration README.

Next step

Separate the objects first

Start here if Source, Memory, and PreparedContext still blur together.

Read the mental model

Start a local Server

If the objects are already clear, write one Memory, find it, and get a cited PreparedContext.

Run the first loop