You've re-explained your project setup to your AI assistant for the last time — or so you thought, until you opened a new session and it forgot everything again.
The question developers keep asking: can AI actually remember how I work?
Short answer: yes, but only certain tools, and only at certain levels.
Three levels of AI memory
Level 1: Conversation memory
What most chatbots offer. The AI remembers what you said earlier in the current conversation. Close the window, gone.
Useful for multi-turn conversations. Useless across sessions.
Level 2: Project-level memory
Config files that persist across sessions — CLAUDE.md, .cursorrules, project-specific instructions. You write and maintain them manually.
Useful for team conventions and architecture decisions. But the AI doesn't update these based on what it observes. You are the memory system.
Level 3: Behavioral memory
The AI observes your actual work sessions — screen recordings, tool usage, coding patterns — and builds a persistent model of your behavior.
Useful for everything the other levels cover, plus patterns you never explicitly documented. The AI notices that you always create a feature branch before writing code, that you run tests before pushing, that you check CI after every push, that you update the changelog for user-facing changes.
Now imagine it notices when you skip a step: "You usually run tests before pushing. Skip this time?"
That's behavioral memory. It knows your process, not just your project.
Why this is hard
Workflows are implicit. You don't write down that you check Slack after deploying. You just do it. The AI has to observe and infer.
Workflows vary. Your hotfix process differs from your feature release process. The AI needs to distinguish contexts.
Workflows evolve. What you did in January isn't what you do in April. The model needs to update without losing useful history.
Privacy is non-negotiable. Workflow recordings contain API keys in terminals, credentials in browsers, proprietary code. The memory system must handle this carefully.
What you can do today
While waiting for perfect behavioral memory:
- Write explicit context files. Start with 10 rules, not 50. Cover conventions, not preferences.
- Use tools with persistent context. Choose assistants that maintain state across sessions.
- Record your workflows. Even if the AI doesn't learn from them automatically yet, recorded sessions create a searchable knowledge base.
- Adopt tools building toward Level 3. The tools investing in behavioral learning now will have the richest models in 6–12 months.
What we're building
Distill builds Level 3 memory through session recording. Each work session adds to your personal skill library. Over time, the system identifies patterns, proposes rules, and detects drift. All processing happens locally — your workflow data never leaves your Mac.
The right question isn't "can AI remember?" It's "does it remember the things that matter — my process, my patterns, my judgment calls — not just my project settings?"