Personal memory

Separately from indexing pages, Trace maintains a personal memory store: a set of natural-language facts about you, extracted automatically as you browse and chat. This is what lets the AI sidebar skip the "let me explain my background" step every time.

Where facts come from

• Social profiles: LinkedIn, GitHub, Twitter/X, Bluesky, ResearchGate, Academia.edu, Behance, Dribbble, Medium, Dev.to, About.me, and Stack Overflow are all recognized sources, each with its own confidence level (LinkedIn and GitHub are trusted the most).
• Browsed pages: Trace infers facts from what you read. Reading several articles about rent disputes might infer "likely a renter," for example.
• Conversations: things you say in the AI chat are mined for facts too. Mention you're launching a startup next month, and that becomes a stored fact.
• Uploaded documents: PDF, TXT, and Markdown files (a CV, for instance) are parsed with high confidence.
• Manual entry: you can always add, edit, or delete a fact directly from the Profile panel.

How Trace keeps the list clean

A few mechanisms stop your profile from turning into a pile of duplicate or contradictory statements:

• Add, update, delete: the model sees your existing memories before extracting new ones, and explicitly updates or removes outdated facts instead of just appending forever.
• Deduplication: exact repeats are collapsed (with a mention count that goes up each time), and near-duplicates are caught by comparing semantic similarity. Anything above 0.85 similarity to an existing fact is treated as the same fact.
• Mention-count gating: an inferred fact (as opposed to one you typed yourself or that came from a high-confidence source) has to be observed at least twice before it's used in AI context. This keeps one-off noise out of your profile.
• Clean up: a manual button (and an automatic pass once you have 10+ inferred facts) runs an LLM merge, one category at a time, that resolves near-duplicates, contradictions, and miscategorized facts into clean, canonical statements, and drops generic ones that carry no real signal.

Where facts show up

Every AI chat turn gets a compact summary of your high-confidence facts injected automatically, for example "works as a designer, lives in Berlin, interested in sustainable architecture." You never have to repeat this yourself.

When Trace extracts new facts, it also searches your existing memories for the most topically relevant ones and includes those in the extraction step, so related facts stay consistent with each other over time.

The graph view

Open the Profile panel to see a canvas-based graph: a "YOU" node at the center, connected to fact nodes colored by category, each linked to the pages they were extracted from. You can filter the graph by keyword to find a specific thread.

Facts are grouped into categories: identity, contact, work, location, life situation, interest, skill, relationship, and general.

Note:

Everything above runs on whichever AI provider you've configured. If you're on Ollama with fully local AI, fact extraction never leaves your machine either. See Choosing your AI.