ianb.rent/ Claude memory
{{ ctaQ }}
Better than Claude’s memory today?
{{ popKicker }}

Glad you think so.

I'm applying for Product Manager, Claude (App Foundations). If you work at Anthropic, a referral is the biggest help, and it takes two minutes:

  1. In Greenhouse, open My Dashboard → Add a Referral and pick Product Manager, Claude (App Foundations).
  2. Paste in my details. One click copies them all.
Even better: I haven't applied yet. Send me your Greenhouse referral link (My Dashboard → Add a Referral → copy your link) and I'll apply through it, so the referral is credited to you.Send Ian your referral link ↗
{{ kitText }}
The job posting ↗ · ianb.rent ↗ · LinkedIn ↗

Fair. Tell me why.

If something here is worse than Claude's memory today, wrong, or missing, I'd love to hear it.

Opens your email app, addressed to me.

What is this?

I'm Ian Brent. I just came off a two-year bender in the AI weeds, building with Claude Code and Codex all day, every day. I have a product background and strong opinions on memory, as you can see on this page and in my writing ↗.

This page is a demo of the memory I think Claude should have. The people are fictional, and the answers are illustrations.

I'm applying for Product Manager, Claude (App Foundations). If you work at Anthropic, please refer me. Here's everything you need →

Think this is a downgrade? Tell me why →
Today · Jordan

Claude's memory today

A set of topics Claude saves as you chat, or when you say “remember this”. You can read, edit or delete each one in Settings › Memory.

Settings › Memory14 topics
{{ tp.t }} EditDelete
Search past chats Paid plans · runs as a tool call
ProjectHalf-marathon plan · its own memory
SensitiveHealth topics aren't saved unless you opt in

Per Claude's help center: editable topics plus past-chat search. It documents no record of when a fact stopped being true. Same conversations as the graph, written as topics, in the words they were said in. Nothing documented connects “his partner” to Sam or “his mother-in-law” to Helen. What a topic list doesn't hold is the rest: the inferences, the links between people, the history of what changed, and the forecast. Linda's knee surgery is a health topic, so it isn't saved by default.

Should have · Jordan

The memory Claude should have

The same person as a knowledge graph. Edges name the relationship; nodes hold the values and their dates. Drag, pinch or ⌘-scroll to zoom, click a node for its source.

{{ heroCard.typeLabel }}
{{ heroCard.label }}
{{ heroCard.stamp }}
{{ heroCard.head }}{{ heroCard.src }}
{{ heroCard.endHead }}{{ heroCard.whyEnd }}
One record type · shape is just a label: Person Organization Place Interest Preference
Stored differently: Event · one date Key date · repeats Commitment · open until done Inferred fact · confidence + reasoning Pattern · derived from evidence World · outside source
Relationship that ended · kept, dated 11–5 · the five differences below
Jordan, Maya, Theo and Priya are fictional See it in action →
What is different

Five things a list of saved topics can't do

Each card is one message. “Claude today” shows what a list of saved topics plus past-chat search has to work with; the answers are illustrations, not recorded outputs. The lit nodes and edges are the part of the graph that made the difference.

{{ L1.point }}

{{ L1.stepNextLabel }}
Maya · {{ L1.s.tab }}
{{ L1.s.msg }}
{{ L1.s.draftHead }}

{{ L1.s.draft }}

Claude: “{{ L1.s.note }}”
Dana’s reply, pasted in by Maya
“{{ L1.s.reply }}”Maya: “{{ L1.s.aside }}”
{{ L1.s.update }}
Claude today

{{ L1.s.plain }}

Claude with this memory

{{ L1.s.mem }}

{{ L1.s.used }}
Claude, to Maya

{{ L1.s.toMaya }}

{{ L1.s.toMayaNote }}
Confidence bands
{{ b.r }}{{ b.l }}
{{ L1.accuracy }}
{{ c.who }} · {{ c.when }}
{{ c.q }}
Claude today

{{ c.plain }}

{{ c.has }}
Claude with this memory

{{ c.mem }}

{{ c.note }}
{{ kd.n }}{{ kd.d }}
{{ pt.t }}
Deliberately not used
×
{{ x.l }}{{ x.r }}
{{ x.w }}
Lit: what this answer used
Three personas · three acts

See it in action: three lives, one year each

{{ nextLabel }}
Hello my name is
{{ P.name }}
{{ P.meta }} · “{{ P.tag }}”
  1. {{ f.n }} {{ f.t }}
{{ yearLabel }} · drag to scrub
{{ k2Label }}
{{ m.m }}
{{ P.healthNote }}
{{ P.name }}'s messages
{{ it.date }}
{{ it.msg }}
{{ it.noMsgText }}
{{ it.change }} {{ it.sum }}
{{ latest.label }} {{ latest.sum }}
{{ p2Card.typeLabel }}
{{ p2Card.label }}
{{ p2Card.stamp }}
{{ p2Card.head }}{{ p2Card.src }}
{{ p2Card.endHead }}{{ p2Card.whyEnd }}
{{ P.name }} · {{ M.when }} “{{ M.q }}”
Claude today

{{ M.plain }}

{{ M.has }}
Claude with this memory

{{ M.mem }}

{{ M.note }}
Used
→
{{ u.l }}{{ u.w }}
Deliberately not used
×
{{ x.l }}{{ x.r }}
{{ x.w }}
Lit: what this answer used · ×: deliberately not used · point at a line to find it
{{ P.healthNote }}
{{ nextLabel }}
What it stores, and how
ThingsPeople, organizations, places, interests and preferences. One record type; the label is only for reading. Different names for the same one resolve to one node: “my partner” is Sam, “my mother-in-law” is Helen.
RelationshipsA labeled line between two things, with the dates it was true. When it ends, it stays in past tense until you delete it, and it can become current again.
EventsSomething that happened or will happen, on one date.
Key datesA date that repeats, like a birthday, even when it’s only implied: “Dad would’ve been 60 today” makes Feb 20 a key date. Whether Claude brings it up depends on the moment.
Inferred factsSomething you never said outright, stored with its reasoning, its evidence and a confidence that gets confirmed or dropped over time.
PatternsDerived, never stated: linked to every event behind it, with a count, and recomputed as evidence comes in.
CommitmentsSomething you said you’d do, kept open until it’s done or dropped.
The worldWeather, news and events from outside sources, with the source and the date they describe. Looked up only for what’s already in your graph, close to when it matters.
Design rules
Your words, pasted material, outside sources and inferences stay labeled as what they are. Something Claude suggested isn't a fact about you until you confirm it.
Names and places stay in the fact ("parents in California", never "back home").
Relative dates become real dates, resolved from when you said them ("next Wednesday" on Nov 3 becomes Nov 11). If a date is ambiguous, it asks.
Every fact keeps its source and when it was last confirmed. Going unconfirmed lowers its weight; it doesn't make it false.
How I'd test it

This page is a demo with scripted examples, not a benchmark. To test it, I'd run saved topics plus past-chat search and this graph on the same conversations, then measure stale answers, unsupported inferences, reminders nobody wanted, and how long it takes to fix a wrong memory.