GitLearnOS

Learner-owned Git memory · GitLearnOS-exclusive native DeepSeek Harness support

Open the core-ready Quickstart →

中文 ·
Website ·
Documentation map ·
Protocol

The core promise

GitLearnOS gives one capable, replaceable main AI agent a learner-owned Git
memory. It notices useful learning events, connects evidence to a goal, guides
the next action, and leaves an inspectable, reversible record.

Learning can happen with a teacher, in class, on paper, in a book, on a
practice platform, in a project, with peers, or with another AI. GitLearnOS
does not move all learning into one application. The main agent connects only
the evidence that is useful for the learner's next decision.

The core is ready with a private local Git repository, one write-capable agent,
the GitLearnOS protocol, and one subject or real learning event. GitHub and
other remotes are optional. During normal learning, Git stays in the
background: the learner should not have to manage folders, branches, or
commits.

The default recommendation is private. Add a remote only when the learner
chooses backup, cross-device continuity, teacher review, collaboration, or
publishing, and keep private answers and gaps separate from shared material.

Core-ready, then everyday learning

The agent answers the immediate request first. Under safe-auto, it may then
make the smallest safe writeback when the target, evidence, goal, and privacy
boundary are clear. preview proposes the exact change without writing;
manual waits for approval. Original answers, notes, and external feedback
are preserved; AI interpretations remain revisable and must link to evidence.

Once configured, a learner should not need to say “use GitLearnOS” or name a
Skill. A question, attempted answer, page photo, class note, teacher comment,
practice result, or repeated difficulty can be a learning event. Incidental
conversation is not stored.

goal and real input
→ organize traceable evidence
→ ask from the current gap
→ keep the answer or external feedback
→ recheck independently later
→ update state with one reversible Git commit

The success condition is better independent performance on later questions,
not a larger pile of generated notes. A normal receipt distinguishes what the
current agent did immediately from what a verified scheduled run actually did.

A small, useful repository

gitlearnos.yml
AGENTS.md
automation.md
dashboard.md
learner-profile.md
subjects/
└── <subject>/
    ├── goals/
    ├── sources/
    ├── models/
    ├── knowledge-gaps/
    ├── reviews/
    └── events/

Only create folders when real learning evidence needs them. Root files hold
shared configuration and current views; subject folders hold focused state. Large
textbooks, PDFs, scans, media, and long-lived references belong in Project
Sources or an authorized local folder. Git keeps compact state, provenance,
selected excerpts, and history.

GitLearnOS-exclusive native DeepSeek Harness surface

GitLearnOS ships an installable native bundle for the official DeepSeek Harness
Developer Preview. It brings a complete, verifiable Git learning transaction
and an agent-controlled panel into Harness: the main agent owns the ordered
Next up queue and presentation decision, while the learner keeps the final
manual toggle. The Host remains bounded plumbing; it does not invent a ranking
or turn panel state into learning evidence.

The code in this repository currently proves:

• a no-build Host plus browser bundle discovered by the Harness profile;
• learning_status and learning_route bounded, read-only observations;
• one gitlearnos.yml-authorized learning_apply transaction that atomically applies typed
  event, knowledge-gap, model, review, and dashboard operations in one
  reversible Git commit (with strict learner identity, setup/config, base
  revision, and write-authority checks); learning_record remains a
  compatibility wrapper;
• a loopback-only, read-only panel that shows the agent-maintained queue,
  respects Panel: expand|collapse, and labels development sample data;
• five queue actions that place a review, practice, close-with-one-question,
  ask-a-teacher, or read-notes request into the conversation input.

RAG provider access remains an optional separate layer; it is not built into
this bundle. DeepSeek's default provider is text-only, but Harness itself is not
limited to text. The learner may either configure a third-party multimodal model
with image input or keep DeepSeek as the main model and install an authorized
vision/OCR bridge plugin. Without either, the agent asks for a transcription
instead of guessing. An immediate multiple-choice answer is supported evidence,
not proof of mastery. Recurring checks use a real scheduler to wake the same main
agent; they do not require a second learning agent. See the
launch note and the
adapter's limits and verification steps.

RAG and background work are separate layers

For substantial textbooks, course packs, notes, or durable personal knowledge,
we recommend (but do not require) a local RAG layer. RAG-Anything
is the first explicitly supported option, not a lock-in.

• Git is the formal, readable source of learning truth.
• RAG is a rebuildable retrieval layer for authorized sources and promoted
  durable knowledge; it is managed by the same main agent, not a second agent.
• The current agent can organize evidence and commit an immediate change.
• Scheduled automation asks a real repository-capable scheduler to wake the
  same main agent. A date, reminder, Harness session schedule, or requested
  marker is not proof that a run happened. maintenance and due-review are
  complete only after each recurring task is created and observed in a real
  scheduler.

RAG may be declined and GitLearnOS still works. One-off exercises do not enter
RAG automatically. If the main agent already understands an image, preserve a
faithful Markdown or structured representation instead of repeating OCR; a
text-only agent must not infer unseen visual content.

Start with one subject

Use the core-ready Quickstart, which contains the single
canonical setup prompt. It asks the agent to identify the private target,
confirm the learning goal, subject, and current material, recommend local RAG,
wait before learner deployment, detect actual capabilities, and report the
undo boundary. The website CTA links to that same source; it does not create a
repository or pretend that a button provisioned a scheduler.

The AceSAT demo shows the loop with a fictional learner using
short, text-first interactions. It still requires a capable AI runtime; local
Git is not the same as a completely offline AI system. Fully offline use would
also require a local model and local tooling that the current runtime actually
supports.

See the impact statement and the completed
SAT fixture for the evidence behind the demo.

Truth before completeness

• Original evidence is preserved; corrections are linked records, not silent rewrites.
• Important conclusions point to traceable evidence; missing evidence stays unknown.
• External resolution and delayed independent mastery remain separate.
• A dashboard is a current view, never a second source of truth.
• GitLearnOS never claims a write, commit, RAG retrieval, scheduler run, Skill
  installation, or mastery without direct evidence.

See GITLEARNOS.md for the full behavior contract,
QUICKSTART.md for deployment, and
Evaluation for documented end-to-end scenarios.

Project status

This branch develops the Git-native v2 protocol and the DeepSeek Harness
Developer Preview. MIT License; see LICENSE.