ReMyndai for consultants, and the client work problem
consulting is the purest version of a job that runs on context you are not allowed to keep in one place.
three engagements, three sets of confidential material, three sets of people, and a brain that has to switch between them cleanly. then a fourth client asks why you recommended the thing you recommended eight months ago, and the answer was in a call nobody recorded.
generic ai tools handle none of this well, and the reasons are structural rather than fixable by a better model. here is what actually goes wrong, and how to set up a memory you could defend in a review, as of september 2026.
why generic assistants fail on client work
they are built around one continuous person, and consulting is not that shape.
every provider shipped memory in 2026 and all of it works the same way: one pool per account, accumulating across everything you do. openai launched a background memory system on june 4, 2026, anthropic reworked claude's into editable entries on july 10, 2026, google rebranded gemini's context layer as personal intelligence in january 2026. we compared them in ai memory is everywhere in 2026.
that design is right for an individual and wrong for an engagement based practice.
| what consulting needs | what provider memory gives you |
|---|---|
| isolation between clients | one shared pool |
| retention tied to the engagement | retention tied to your account |
| deletion on client request | deletion of your whole memory, or nothing |
| an archive you can produce in a review | an archive the vendor controls |
| no training on client material | depends on the plan and the terms |
the first row is the one that matters. cross contamination is the failure that ends relationships, and a shared memory makes it structurally easy rather than merely possible.
the confidentiality questions worth answering first
client confidentiality is settled by three questions, and they are contractual before they are technical.
what does the engagement actually permit? the msa or nda usually addresses disclosure to third parties, and a cloud model is a third party. nda compliance here is usually a question of which tier and which account, not whether ai is allowed at all: many enterprise agreements permit an approved tool list and forbid consumer accounts.
does the vendor train on your inputs? "we do not sell your data" and "we do not train on your data" are different sentences and you want both, explicitly, in the tier you are actually paying for.
can you delete on request? if a client asks you to purge their material, can you do it without destroying everything else? if the answer is no, that is a design problem you have already accepted.
there is a fourth question people skip, which is what happens if the vendor is acquired. rewind spent 2 years telling users their archive was theirs, and in december 2025 meta acquired the company and disabled all capture on december 19, 2025 with 14 days of notice. the general pattern is in what happens to your data when an ai app shuts down.
where the hours actually go
context switching, and it is invisible on a timesheet even though it eats billable hours.
the expensive part of consulting is not producing the deliverable. it is rebuilding the state of mind that produced it. you leave engagement a on thursday, spend two days on b, come back monday and spend 40 minutes remembering where you were, what the client had objected to, and why the third option was ruled out.
none of that is in the deck. decks record conclusions and discard reasoning, deliberately, because that is what a deck is for. so the reasoning lives in a call nobody recorded, a document read once, a thread that scrolled away, and your memory, which is doing its best.
per the anthropic economic index for may 2026, the number one work task in sampled ai conversations is searching electronic sources for information at 4.95%, with reference searching second at 3.74%. retrieval is the dominant activity. now look at what is not retrievable.
| request topic | share of conversations |
|---|---|
| content creation and copywriting | 22.72% |
| education and learning | 13.23% |
| software development | 11.51% |
| research and intelligence | 10.94% |
| document processing and extraction | 4.32% |
| knowledge retrieval and enterprise search | 3.61% |
| personal ai assistant | 2.86% |
| conversation and meeting intelligence | 0.26% |
searching your own organisation's knowledge is 3.61% and anything about a meeting is 0.26%. for a practice whose entire product is judgement formed in meetings, that second number is the whole problem.
knowledge management for consultants, honestly
the firm wiki failed for the same reason your personal one did.
every practice has tried this. a shared drive, a wiki, a lessons learned template nobody fills in after the third engagement.
microsoft and apple both attacked the same gap from the platform side in 2026, with recall on windows and siri ai's onscreen awareness and personal context on june 8, 2026, assessed in is microsoft recall safe and apple intelligence on mac. neither is scoped to an engagement, which is the thing a practice actually needs. the reason it decays is not discipline, it is that filing costs a real interruption each time and the payoff arrives months later for someone else. we went through the maintenance problem in the best second brain apps in 2026.
what survives is what records without being asked. that is why meeting transcription stuck where wikis did not: it costs nothing per item. the limit is that it only holds the audio channel, so the number said out loud makes the transcript and the slide it was written on does not, which we covered in what ai note takers miss.
setting up an archive you could defend in a review
six things, in order, and they take about 15 minutes once.
- own the machine, or get it in writing. the cheapest protection available, and the first question anyone will ask.
- exclude before you capture. anything belonging to a client who has not agreed, plus password managers, banking, hr and health portals. a tool that cannot exclude by app and by site does not belong on this laptop.
- keep the archive local. an archive that never leaves your disk is a far smaller problem under every regime, and it means no third party is holding your clients' material on your behalf.
- set a retention limit and honour it. "forever" is the hardest position to defend. tie it to your engagement records if you can.
- default the audio off. turn it on per call, after announcing it. two party consent statutes govern conversations rather than pixels, so the audio toggle is the one carrying legal weight, as set out in is it legal to record your screen at work.
- be able to delete on request. if you cannot demonstrate this, do not run the archive.
none of that is exotic. it is the same posture you already apply to files, applied to a category people have not thought about yet.
where remynd fits, and where it does not
it is the archive layer, on your own mac, and the scope is worth stating exactly.
remynd captures the focused window rather than every display, runs ocr locally so what you looked at becomes searchable text, and keeps the index on your machine. call transcription is on device too, running an mlx speech model on apple silicon. you can exclude specific apps or sites from capture entirely, retention is a setting that defaults to 30 days of recordings, and your history is read only everywhere in the app.
what it does not do: asking a question sends the retrieved slices to a cloud model, and there is no local model option today. it also does not isolate engagements into separate vaults, so the exclusion list is doing that work rather than the architecture. if strict per client isolation is a contractual requirement for you, that is a real gap and you should know it before installing. the fuller architecture is in private ai on your mac, and the developer view is on for engineers.
the honest framing
the value is not billing more hours. it is answering the question you cannot currently answer.
"why did we recommend that" is the question that turns a project into a relationship, and it is the one nobody can answer eight months later because the reasoning was never written down anywhere retrievable. that is a recording gap rather than a memory failing, which is the argument in why you forget what you read.
fix the recording and the rest follows. just set the exclusions first, because on client work the thing you did not capture is worth more than the thing you did.
download remynd for mac and set your exclusions before your first client call.