Claw-ED: The Lesson Planner That Remembers What You Taught Last Week

Most lesson-planning AI tools behave like fast, forgetful interns. You give them a topic, they give you a draft, and the next prompt starts from zero. Claw-ED is trying to do something harder: remember what the class already covered, what the teacher prefers, and what should logically come next.

That matters because good teaching is sequential. A lesson on the French Revolution lands differently if the system knows the class already covered the Enlightenment, reviewed taxation, and used a particular scaffold last week. Claw-ED’s pitch is that a lesson planner should understand that timeline, not just the topic.

The Best AI Lesson Tool Is the One That Remembers Your Classroom

Claw-ED is built around that idea. It is not just an LLM wrapper that spits out worksheets. It is a local-first AI co-teacher that ingests a teacher’s materials, learns the recurring shape of their instruction, and uses prior classroom context to guide the next output.

That shifts the product from content generation to continuity. Instead of asking, “What should a lesson on this topic look like?”, Claw-ED asks, “What have you already taught, how did you teach it, and what does the next step need to preserve?”

Why a Teacher’s Voice Is More Than Word Choice

The repository’s most interesting idea is not style matching in the superficial sense. It is trying to learn a teacher’s pedagogical fingerprint: the way they sequence explanation, where they put checks for understanding, how they scaffold, and how much they expect students to carry on their own.

That is what the project’s self_distill.py and “soul” concept are aiming at. Strip out the generic AI habits. Preserve the teacher’s structure. If the output sounds like the teacher, that is because the system has learned more than vocabulary. It has learned the shape of instruction.

Inside the Memory Stack: Identity, Curriculum State, Episodic Recall, and a Knowledge Graph

Claw-ED does not rely on one memory store. It assembles context from several layers: teacher identity, curriculum state, episodic recall, session summaries, and a temporal knowledge graph. That is the difference between searching your files and understanding where you are in the school year.

That matters because the graph changes what counts as relevant. A topic is not just semantically related to another topic. It may be the thing the class reviewed yesterday, the prerequisite it still struggles with, or the bridge to the next unit. In Claw-ED, memory is about sequence, not only similarity.

The Privacy Bet: Local Embeddings, SQLite, and No Need to Phone Home

Privacy is not a side benefit here. Teachers handle student work, classroom notes, and planning material that often should not leave the device. Claw-ED’s local-first approach, including ONNX embeddings and SQLite-backed state, is a product decision shaped by that reality.

The trade-off is clear. Local execution is slower to set up than a hosted SaaS tool, but it reduces trust friction. For a school district or a cautious teacher, that can matter more than a polished onboarding flow.

How Claw-ED Turns One Request Into a Full Teaching Package

The project’s agent loop is what turns that memory into action. A request like “plan my week” does not produce one text blob. It can trigger planning mode, tool iteration, and a chain of discrete outputs that form a 9-file bundle: lesson plans, handouts, slides, assessments, and related classroom materials.

That bundle is the delivery mechanism, not the differentiator. The differentiator is that each file is meant to feel like it belongs to the same teacher, the same unit, and the same moment in the curriculum. The loop, planner, and tool registry are there to keep that coherence intact.

# Conceptual sketch of the agent loop
for step in range(MAX_TOOL_ITERATIONS):
    action = llm.decide(next_context)
    result = tools.run(action)
    memory.update(result)
    if result.done:
        break

# The planner can switch the agent into multi-step mode
# Search -> Generate -> Export

That is the practical value of the agent design. It is not trying to be clever in one shot. It is trying to be consistent across steps.

What It Replaces, and What It Doesn’t

Claw-ED is not a replacement for a teacher. It is a replacement for the repetitive part of lesson prep, especially when the same standards, scaffolds, and formatting rules keep reappearing. Compared with ChatGPT or Claude used directly, it adds memory, workflow, and output structure. Compared with generic edtech tools, it adds local control and model flexibility.

That is why the project feels more like an operating system for lesson prep than a chatbot. It tries to sit inside a teacher’s existing workflow, not outside it.

The Trade-Offs: Power, Complexity, and a Lot of Moving Parts

The ambition is real, and so is the complexity. A system that learns voice, preserves sequence, and runs locally is doing several hard things at once. If it works, the teacher gets continuity. If it breaks, the stack has many places to fail.

That is the right tension for this kind of project. Claw-ED is interesting because it refuses the shallow version of educational AI. It does not just automate content. It tries to remember a classroom.