🚀 Resume Feedback Is Easy — Until You Try Making It Context-Aware

While building my AI Career Advisor, I initially assumed resume feedback would be one of the simplest features.

Upload resume → analyze → give suggestions.

There are already dozens of tools doing this, so it seemed straightforward.

But once I introduced memory and user context, things became much more complex.

🧠 What the system actually needs to do

At a basic level, the system should:

• Parse a resume
• Extract skills and projects
• Compare them with target roles
• Suggest improvements

Simple enough — but in practice, this wasn’t sufficient.

Because a resume is only a "snapshot", not the full story.

⚠️ The real problem: resumes are incomplete

Users often:

• Forget to include recent work
• Undersell their projects
• Omit important details

If the system only analyzes the uploaded resume, it misses critical context.

So the real challenge became:

👉 'How do we combine resume data with stored user memory?'

❌ First attempt: treat resume as the source of truth

const parsedResume = parseResume(file);

const response = await llm.generate({
  input: "Give resume feedback",
  context: parsedResume
});

This worked — but only at a surface level.

It couldn’t detect missing information or inconsistencies.

✅ The fix: merge resume with memory

const memory = await hindsight.retrieve(userId);
const resumeData = parseResume(file);

const context = {
  resume: resumeData,
  pastProjects: memory.projects,
  skills: memory.skills
};

Now the model has access to:

• What the user wrote
• What the system already knows

💡 Why this matters

This enables insights like:

• “You worked on X but didn’t include it”
• “Your project description is too vague compared to stored details”

This kind of feedback is impossible without memory.

🔗 Using Hindsight for context

The memory layer is powered by:
👉 https://github.com/vectorize-io/hindsight

More details:
👉 https://hindsight.vectorize.io/

Concepts:
👉 https://vectorize.io/features/agent-memory

📈 What improved after this change

Before:

• Generic resume tips
• Repetitive suggestions

After:

• Context-aware corrections
• Missing content detection
• Better personalization

❌ What didn’t work

• Relying only on resume text
• Ignoring past interactions
• Overloading the model with full memory

🧩 Lessons learned

• A resume is incomplete without history
• Memory enables comparison, not just analysis
• Context merging is more powerful than parsing

🏁 Final thought

Resume feedback is easy to build.

Context-aware resume feedback is not.

And the difference is "memory".