Use litecode

Memory and learning

litecode saves every session to disk. It gets better with use in three ways. None of them change the model weights.

What is saved, and where#

Data	File	Scope
Every turn of every session	`<project>/.litecode/sessions/<id>.jsonl`	project
Full history before compaction	`sessions/<id>.full.jsonl.bak`	project
Summary of older turns	`summaries/<id>.md`	project
Project facts	`memory.jsonl`	project
File snapshots for undo	`checkpoints/`	project
Security review reports	`reviews/`	project
Your preferences	`<data_dir>/profile.jsonl`	all projects
Finished tasks + ratings	`<data_dir>/learn/examples.jsonl`	all projects

If project_memory = "central", the project files go to <data_dir>/projects/<name>-<id>/ instead of the project folder.

All files are plain JSON lines or Markdown. You can read, edit or delete them.

Preferences. A message line that starts with "always", "never", "don't", "prefer", "from now on", "make sure to" or "avoid" becomes a preference. No model call is needed. Preferences go into every prompt, in all projects. Add one by hand: /prefer use pathlib, not os.path.
Facts. A test command that passes is saved at once. When a session ends (/exit), the model reads the session and saves up to 6 durable facts. To turn off the model step: litecode config auto_learn false.
Examples. Each finished task is saved with the files it changed. For a new task, litecode finds similar past tasks by keyword overlap. It adds the best one to the prompt. Rate answers with /good and /bad. A bad answer is never reused. A good answer ranks higher.

Context limits: small models have small context windows. litecode keeps the last 8 turns raw. It folds older turns into a summary.

Constraint: training a 2B+ model on a 4-core CPU takes days. Do it on a GPU machine. A free Google Colab GPU is enough for a LoRA on a 2B-4B model.

Rate answers with /good while you work.
Export: litecode learn export my_data.jsonl (add --all to include unrated tasks).
Train a LoRA on the same base model you use (for example with Unsloth or PEFT). The file is chat format: {"messages": [...]} per line.
Convert the adapter to GGUF with llama.cpp: python convert_lora_to_gguf.py --base <base model dir> <adapter dir>.
Copy the .gguf adapter to your PC.
Set it: litecode config lora D:/llm/adapters/my-lora.gguf.
Restart litecode. The server loads the adapter.

Warning: an adapter works only with the base model it was trained on.

Nothing leaves your PC. litecode makes no network calls except litecode setup downloads, and MCP servers that you add yourself.