Ask people what worries them about AI tools and you'll hear about wrong answers, bias, job displacement. Ask what should worry a person typing their private struggles into one, and the honest answer is quieter: memory.
A stateless chatbot — one that forgets you between conversations — can only ever mishandle the message in front of it. A stateful one is different in kind, not degree. Session by session, it's accumulating a picture of you: your patterns, your relationships, your recurring 2 a.m. subjects. That accumulation is exactly what makes an AI coach useful — advice from something that remembers your context beats advice from something meeting you for the first time. And it's exactly what makes one dangerous, because "a tool that remembers your context" and "a growing file about your inner life" are the same object described in two tones.
Memory is the highest-stakes design choice in an AI coach. Every tool in this category made a decision about that trade — a decision you'll rarely find written down. This post tells you ours, in the form of a remembers/forgets inventory verified against our own system, and then turns the inventory into the questions worth asking any tool that holds your thoughts.
Three data paths, one sentence each
First, a map in words, because "what does the AI do with my data" is actually three separate questions that deserve three separate answers.
The outbound path — what leaves for the AI provider when you send a message — is the redaction story: identifying details are stripped before any text reaches an external model. The training path — whether your content teaches future models — is the training story: your entries aren't training data. The retention path — what the system keeps about you afterward, and recalls next session — is this post. Three different pipes. A tool can be honest about one and silent about the others, which is why the questions at the end come as a set.
What ours remembers
The design choice we made, stated before the details: our coach's long-term memory is summaries, not transcripts. It doesn't re-read your old conversations. It reads short, deliberately blurred summaries of them. To be precise about what that does and doesn't mean: your conversation history itself is stored — encrypted, in your account, yours to reopen and search anytime. The design choice is about the coach's memory specifically: what it carries forward between sessions is never the transcript. Everything below follows from that one decision.
Recent conversation, verbatim — briefly. Within a session, the coach sees your recent messages as you wrote them (redacted, per the outbound path). That's what makes conversation possible at all. But the window is deliberately budgeted: as a session grows, older turns fall out of the live context rather than accumulating forever.
A rolling session summary. What carries forward isn't the transcript — it's a running summary the system maintains as the conversation grows, so the coach can remember that you've been working on the job decision without re-reading every word you wrote about it. The summaries are built under strict rules: identifying details are stripped before summarizing, the output is written in deliberately generic language — pronoun-neutral, names and places generalized — and the stored result is scrubbed a second time by the same redaction machinery that guards the outbound path. Belt, then suspenders. What lands in storage is closer to a case note than a diary: "the user has been weighing a career change; sleep is a recurring obstacle."
Summaries of your goals and journal context. The same summarize-then-scrub treatment extends to what the coach knows about your goals and recent journaling — so its advice can connect to your actual life without your actual sentences being hauled into every prompt. Two of the taps are yours to close: whether the coach may draw on your goals, and whether it may draw on your journal — each has its own switch in your profile settings, and there's a per-coach context override besides. The rest of its working context (your habits, tasks, and assessment picture) currently rides along without a dedicated switch — a gap we're naming here because this post exists to name exactly such things.
All of it, encrypted. The stored summaries — like your raw entries themselves — are encrypted at rest. A database backup of our system does not contain readable copies of the coach's notes about you.
One message's life, step by step
To make the pipes concrete, follow a single Tuesday message through the system: "Talked to my manager Karen again about the transfer. Slept terribly worrying about it."
In the live conversation, the coach responds to your actual sentence — with identifying details stripped from what any external model sees, so "Karen" doesn't travel. As the session grows and older turns fall out of the live window, the rolling summary absorbs what matters going forward — something like "weighing an internal transfer, a strained dynamic with their manager, sleep disrupted by work stress." That summary gets scrubbed again before storage and encrypted. Months later, when you raise work in a new session, what the coach "remembers" is that case note — not the Tuesday transcript, not the name, not your exact words at your most raw. And on the day you delete your account, the note goes with everything else.
No long-term copy of your sentence ever lived in the coach's memory — the encrypted conversation record stayed in your account, for you, while the coach's carried-forward knowledge stayed a blurred note. That's the design in one worked example — and it's also exactly the walk-through to request from any other tool: "trace one message for me." Vendors who can, will.
What ours forgets, on purpose
The forgets column is the more informative one, because forgetting is where a memory design shows its values.
It forgets your exact words, over time. Because long-term memory is summaries, the specific sentences you wrote in a hard moment don't ride along into every future conversation. The blurring is the feature: a coach that needs to know you're working on your marriage doesn't need to carry the verbatim transcript of your worst night, and ours structurally can't quote it back from months ago, because what it retains was never the transcript.
It forgot how to search you in plaintext. A concrete example of the design being enforced after the fact, told plainly because it's the kind of thing this series promised to disclose: our chat search feature originally kept a separate readable copy of message text to search against. We removed it — the column was destroyed, and search now works by decrypting your own messages on the fly, for you, when you ask. The lesson generalizes: memory design isn't a launch-day decision, it's a discipline, and "what readable copies exist?" is a question systems have to keep re-answering.
It forgets you entirely when you leave. Deleting your account is genuinely deletion — entries, summaries, the coach's accumulated context, gone rather than flagged-invisible. The full posture, including how database backups age out on a fixed window, is written up in our portability post; the sentence that matters here is that the coach's memory of you doesn't outlive your decision to go.
And it never remembers into a training set. Retention and training are different pipes, but the fear rhymes: the file about you becoming permanent in someone's model. Per the training post, that pipe is closed.
Why "just forget everything" isn't the safe answer either
It's tempting to conclude that the only trustworthy AI tool is a fully amnesiac one. Worth examining, because the stateless design has quiet costs of its own.
A tool that forgets you completely makes you the memory system — to get situated help, you retell the story, in full, every session, and each retelling is a fresh transmission of raw material. A well-designed memory can reduce total disclosure: context carried as a blurred summary is context you don't re-type verbatim twelve times. Statelessness also can't notice patterns — the recurring subject, the slow slide across weeks — which is much of what makes a reflection tool worth more than a search box.
So the choice isn't memory versus safety. It's which memory design: what's retained, in what form, under whose control. Amnesia is one legitimate answer for one kind of user. A summarized, scrubbed, encrypted, deletable memory is ours. The failure mode is the middle: tools that remember everything, verbatim, indefinitely, and describe it as personalization.
The honest limits
Symmetry demands the other column, and there are real entries in it.
The summaries, though blurred and scrubbed, are still about you. A case note that says "recurring conflict with a sibling; avoids the topic" contains no names and plenty of intimacy. Redaction narrows what memory holds; it doesn't make held memory neutral. If that trade isn't acceptable to you, the honest answer isn't a softer description of our design — it's using the coach with the context taps turned off, or not using one at all. Both are legitimate readings of the same inventory.
Automated scrubbing is strong, not perfect. Unusual phrasings can slip past any redaction rule set — we've said so since the redaction post, and it applies to summaries too. The system is built so that what slips is rare and blurred further by summarization, not so that slippage is impossible. No one can honestly promise the latter.
And memory makes the coach more persuasive, which is its own risk class. A tool that remembers your history speaks with borrowed intimacy, and borrowed intimacy deserves suspicion precisely in proportion to how good it feels. The remembering is in service of reflection, not relationship — ours says so in its design, but the guard that matters is the one you keep.
The questions to ask any tool
Turn the inventory into portable equipment — five questions for any AI tool that holds personal material. A vendor with a real memory design answers all five specifically; a vendor who answers "your data is encrypted" to each of them has answered none.
Transcripts or summaries? Does long-term memory keep what I wrote, or a derived note? For how long is the verbatim text in play?
What's stripped, and when? Is identifying detail removed before the AI provider sees text, before storage, both, or neither?
Can I see and control what it knows? Is there a way to view what the tool carries about me, and to exclude sources — or is memory all-or-nothing?
What does deletion actually delete? The account? The conversations? The derived summaries and embeddings too? "Deleted from view" and "destroyed" are different answers.
Is any of it training data? Separate pipe, separate question, and the answer should be a plain no or a very specific yes.
Our own answers are the four sections above, and the audit-log post covers the companion question — who on the company's side can read what, and what mark it leaves. If a tool you're evaluating can't produce this page's equivalent, that silence is the answer: the trade got made without you.
Memory is what makes an AI coach worth using. It's also the file. Any tool in this category that only tells you the first half is selling; the whole page above is what the second half looks like written down — and the trust page is where we keep the running version.
—
This post describes NexTier's coach-memory design as implemented at the time of writing and was verified against the live system this week; the limits named are real. It is educational content about evaluating AI tools, not a security guarantee, and inherits the fuller disclosures of our redaction and training posts.