Memory Inspector

byo-key · watch what it remembers

Say something about yourself.

Chat normally — I'm a landscape architect in Lisbon, I prefer terse answers, I'm allergic to shellfish. After each reply, Claude extracts the durable facts into the memory store on the right, shaded by how sure and how sensitive each one is. Edit or forget any of them.

Memory

Nothing remembered yet. What the assistant learns about you shows up here after each turn.

Shading

sensitivity →

confidence →

faint = unsurered = private

Your Anthropic API key

This tool calls Claude directly from your browser. Bring your own key.

Your key stays in your browser. It is never sent to any server of ours.

It is stored only in this browser's localStorage and sent straight to api.anthropic.com — nowhere else. This site is static; there is no backend to receive it.
Your chat and the extracted memory also live only in this browser. Nothing is sent anywhere but Anthropic.
Clear it any time with Forget key below.

API key (starts with sk-ant-)

Get a key ↗

How this works

A memory-transparency prototype for whatupwolf.com/lab.

Every turn, the assistant is told what it currently remembers about you (the list on the right), so its replies are shaped by that memory. After it answers, a second extraction pass reads the exchange and decides which durable facts to add or update — and estimates, for each, a confidence (how sure it is the fact is true and lasting) and a sensitivity (how personal or private it is).

Each entry's background shading is a 2-D read: how far along the confidence axis it sits (faint → strong fill) and where it lands on the sensitivity axis (calm green → warning red). What the system "knows", and how sure and how sensitive, is legible without opening anything.
Edit an entry to correct it — an edited fact becomes fully confident (you confirmed it). Forget removes it.
Edits and forgets take effect on the next turn, because the memory list is re-injected into the model's context each time you send a message.
The extraction pass looks for stable, user-specific facts (identity, preferences, goals, constraints) and ignores one-off chatter. It's a heuristic — that's the point of being able to see and correct it.

Inspired by MemoAnalyzer, on making an assistant's inferred memory visible and editable.