Stop paying the retrieval tax.
Every agency owner and solo founder pays a retrieval tax. You made the decision nine months ago. You just cannot find it. It lived half in Slack, half in a call nobody recorded, got confirmed in an email thread and contradicted by a Jira ticket. So you lose three days rebuilding it from fragments, usually on the day it matters most.
Search does not help. Search finds messages. It does not tell you why something was decided, who agreed, or what got ruled out.
We were paying that tax weekly running client work, so we built Zenpa. You ask a question, it pulls the actual sources back together and gives you a decision brief with citations. Share it, or hand it to another AI tool.
It runs local-first, on-device inference. Your client conversations, internal disagreements and unshipped plans are the most sensitive data you own. Uploading all of it to a third party to get answers about your own work never sat right with us.
We just launched. What people keep coming back for: client disputes, meeting prep, and reconstructing project history nobody remembers.
Question for this crowd. How do you actually reconstruct a decision from months ago? Do you have a system, or do you just pay the tax?
Raph
The phrase "retrieval tax" is interesting because it describes a cost most teams accept without ever naming.
Information rarely disappears. The reasoning behind decisions does. Once that context is gone, every future discussion starts from scratch, even if all the evidence still exists.
That distinction is the whole thing. The evidence survives, the reasoning does not.
What we kept seeing is that the reasoning was never written down anywhere in the first place. It lived in the gap between messages. Someone proposes X, someone says "yeah let's do that," and the constraint that made X the right call was in a call two days earlier, or in a message that got resolved and scrolled past. The archive is complete and the answer still is not in it.
Which is why we ended up building a decision graph rather than better search. Search assumes the answer is in a document. Here you have to link people, threads, meetings and timing to reconstruct something nobody ever wrote as a single statement.
And you're right that people accept it silently. Nobody has "reconstruct decision" on a task list. It shows up as a slow week, or a meeting where everyone relitigates something already settled. Costly, but hard to see, which is exactly why it never gets fixed.
Raph
I appreciate you taking the time to unpack that, Raph.
I'd be interested in continuing the conversation by email if you're open to it. What's the best email to reach you on?
Honest answer: we don't.
Zenpa gives you the answer with the sources attached, and the team decides what to do next. Whether a decision still holds is a judgement call we deliberately do not make for you. It's a human problem, and any system is only as good as the information it tracks.
That said, there's a tractable version. A decision has a fingerprint: the people, the constraint, the project, the thing that got ruled out. You could search forward through more recent activity for anything that contradicts it and flag it for review. Not "this is stale," just "here's something from March touching the same constraint, worth a look."
Keeps the judgement with the human, just puts the evidence in front of them earlier. Not built yet, but you're the second person to push on this!
Remembers everything” is powerful, but also broad.
The products that break through usually anchor to a very specific use case first.
What’s the main job people are using it for right now?
Client disputes. Specifically the moment a client says "we never agreed to that."
You just found your anchor — but your homepage is still selling the feature, not the job. The freelancer who hears "remembers everything" thinks "nice for my meeting notes." The agency owner who hears "never lose a client dispute" thinks "this is insurance."
The risk: Broad positioning trains the low-LTV user (meeting notes) to sign up, while the high-LTV user (dispute protection) never realizes you solve their specific pain.
Quick test: Ask new signups "What triggered you to look for this today?" If <30% say "client dispute," your messaging is leaking your best buyer.
I run PreLaunch AI — we simulate buyer segments to find the job that justifies premium pricing vs. the feature that attracts free users. Happy to model your "dispute protection" vs. "memory tool" segments if you want to see where the money is.