You've explained your project structure to ChatGPT for the fifteenth time. Your coding assistant still suggests camelCase when your codebase is all snake_case. Every session starts from zero.
The promise of AI that learns your workflow has been around for years. Only recently have tools started delivering on it.
What "learning" actually means
There's a difference between memory and learning:
- Memory = the AI remembers what you said earlier in this session. Close the window, it's gone.
- Project context = persistent config files (CLAUDE.md, .cursorrules) that you write and maintain manually.
- Behavioral learning = the AI observes your work sessions and builds an improving model of your patterns. This is the one that matters.
True workflow learning requires persistent memory across sessions, pattern recognition that identifies repeatable behavior, actionable output (not just observations), and drift detection when you deviate from established patterns.
What the learning curve looks like
Week 1: Observation. The tool records sessions and builds a baseline. It identifies your most-used tools, common sequences, file organization patterns. Mostly watching.
Week 2: Pattern recognition. After 5–10 sessions, repeatable patterns emerge. You always run tests before committing. You check Slack after every deploy. You organize features in a specific folder structure.
Week 3: Proactive suggestions. The tool suggests next steps based on recognized patterns. It prompts the linting step before you navigate to it yourself.
Week 4+: Compound improvement. The tool proposes rules — formalized patterns it follows automatically unless you override. These become a living document of your working style.
How to evaluate after two weeks
Ask these questions:
- Does the tool suggest things it couldn't have on day 1? If not, it's not learning.
- Has suggestion accuracy improved? Track accept vs. reject rates.
- Can it describe your workflow back to you? A learning tool should produce a summary of your patterns.
- Does it notice when you break pattern? Drift detection is the strongest signal.
If you answer "no" to all four after two weeks of regular use, the tool is marketing "AI learning" without delivering it.
What to watch out for
Confusing memory with learning. An AI that remembers your name isn't the same as one that knows your deploy sequence.
Giving up after three days. Workflow learning needs data. Five full sessions minimum before patterns emerge.
Not reviewing proposed patterns. The AI might identify patterns that are actually bad habits. Review before approving.
Running multiple tools simultaneously. If you're evaluating three workflow tools at once, none gets enough data to learn.
What we're doing about this
Distill takes the session recording approach. It captures your screen and system events, synthesizes each session into a structured brief, and after multiple sessions identifies patterns, proposes rules, and detects drift. The compound loop is the product — not a feature bolted onto something else.
The bar is simple: does it know you better in April than it did in March? If yes, it's working.