The most expensive asset in a cross-border e-commerce operation is not the inventory sitting in a Shenzhen warehouse, nor the ad account you’ve spent six figures learning to run. It’s the undocumented judgment: why the listing rank dropped in March, which supplier delivered late twice, how you finally got the TikTok Shop return rate under control. That knowledge usually lives in a chat thread, a doc nobody updates, or a senior operator’s head. Greplica, the self-updating wiki for coding agents that is making its way through Product Hunt, is not built for sellers. But it is the clearest artifact yet of the operating pattern that will run e-commerce brands in the next two years: continuously extracting tribal knowledge, anchoring it to the artifacts that actually change, and retrieving it only when a decision needs to be made.
The problem: your operating knowledge dies at the end of every session
Every cross-border seller I have worked with has the same stack: a Slack archive that is the only true record of why a listing was rewritten, an Amazon Seller Central account full of old compliance emails, a Shopify store with a checkout flow no one dares touch, a TikTok Shop remarketing setup that exactly one person understands. The knowledge that actually runs the business is not in the SOP doc. It is in the session logs of whoever survived the last Q4.
Developer teams recently discovered the same problem in a harsher form. Coding agents write a growing share of the code, but their memory evaporates at the end of every session. Kushal Patil, co-founder of Greplica, put it directly: “Internal documentation is always boring, but with coding agents now writing a 100% of the code it is more important than ever. CLAUDE .md files can only store so much information - the real secrets of the codebase that your agent learns in its runs get lost in a new session.”
That is what I mean by session amnesia. Replace “codebase” with “catalog operations,” “supplier relationships,” or “Amazon account health,” and the sentence still works. The tools you already run — AI listing generators, chatbot support agents, ad copy assistants — are all losing the most valuable context they produce. Every session is a fresh amnesiac.
What Greplica actually does differently
Greplica is not another documentation tool. It is a shared memory layer for the engineering team and every coding agent that touches the repo. The product description on its launch page says it “continuously extracts decisions, constraints, gotchas, failed approaches, and file-level context from coding sessions, then retrieves only what matters for the task at hand.” Unlike static docs or siloed agent memory, it stays grounded in the repo, keeps knowledge fresh, and works across developers, agents, clones, and forks. It is open source, runs locally, and offers a managed shared mode.
Two details stand out. First, the knowledge is extracted, not authored. Nothing about Greplica asks a human to write better documentation. Second, the facts are anchored in committed code. A fact that points to code that no longer exists has a natural expiry date. One reviewer on the thread said that was the part that convinced him: “it gives stale knowledge a natural expiry date, which hand-written docs never have.”
Compare that with what most teams use today. Notion and Confluence are where documentation goes to rot. GitHub Wikis are versioned but still depend on someone editing them. Claude Code’s CLAUDE.md files are a step toward agent memory, but they are static files with a hard size ceiling. Cursor rules are project-level instruction files, not extracted knowledge. Greplica treats documentation as an output of work, not an input to work. That is a genuinely different mental model.
When someone on the launch thread asked whether Greplica works only with Claude Code or across tools like Cursor, Codex, and Copilot, Patil answered: “Completely agent agnostic.” When another reviewer asked how it sits next to existing CLAUDE.md files, Patil said: “we ingest Claude.md in our system itself. So that you can remove it completely. At least the docs reading part.”
That is the part to pay attention to. It is not trying to be another place where you file documents. It is trying to be the memory that the documents used to stand in for.
Why Amazon sellers should care more than Shopify ones
Shopify sellers can encode operating knowledge into code: custom scripts, app configurations, theme logic, subscription variations. A lot of the tribal knowledge becomes process. Amazon sellers cannot. Amazon Seller Central is a black box. The knowledge that matters — how to handle a policy warning, which supplier to trust with a fragile ASIN, what actually happened during a listing suppression — lives in emails, seller forum threads, and the heads of account managers.
That is exactly the kind of knowledge Greplica’s model is built to preserve. And the cost of losing it is higher. A wrong “we tried X” note in a Shopify ops wiki might cost you a week of ad spend. A wrong note about Amazon compliance can cost you a listing, or an account. For cross-border sellers, the risk profile is not symmetrical. The more black-box the marketplace, the more valuable a self-updating memory of decisions becomes.
What a cross-border operator should steal from this
You do not need to run a dev team to borrow the pattern. Greplica is a useful thing to install on a GitHub repo, but it is more useful as a provocation: what would a living wiki for your whole e-commerce operation look like?
Most sellers use a Notion or Google Doc as a graveyard. You write an SOP in Q1, update it twice, then it rots. Greplica’s trick is to make documentation continuous and retrieval-based. You don’t read the wiki; the wiki reads you.
Here is what I would steal even without touching the tool:
- Separate decisions from definitions. Stop writing “How refunds work.” Write “Why we changed the return window in May, and what happened.” Anchor it to an artifact: the email from the agent, the spreadsheet of return rates, the ad campaign screenshot.
- Give every fact a source. If a fact cannot point to a concrete artifact — an order data export, a support ticket, a campaign snapshot — it should not be trusted.
- Make retrieval part of the workflow. Instead of a human searching the wiki, feed the wiki to an AI assistant that can retrieve relevant context per task. The point is not to organize information. The point is to make information appear at the moment of decision.
- Auto-expire facts. Any claim not touched in 90 days gets flagged. If no one has needed it, it probably does not matter. If it matters but no one updated it, the flag forces a human to decide.
The “SOP wiki” you can build this week
Start small. Pick one area of your operation that has caused the most repeated pain in the last quarter — returns, supplier communication, or listing optimization. Create a decision log with four columns: the decision, the evidence, the date, and the review date. Every time you change something, write one sentence about why. Then wire that log into your AI workflow so that any assistant answering a question about that area is forced to cite the log first.
That won’t be as elegant as Greplica. But it will prove the core thesis: the value is not in the wiki, it is in the extraction and the retrieval.
Where my judgment says the model breaks
The part that actually concerns me is not extraction. It is retirement. Greplica anchors facts in committed code, which handles the “what does this module do” half of documentation. But a large share of the most valuable knowledge in a codebase — and in an e-commerce operation — is negative: “we tried this approach, and it failed.”
Nothing in the code changes when that fact stops being true. A reviewer on the thread, Asad M., articulated the risk better than most launch-day commentary ever does: “One bad conclusion from one bad session becomes ground truth for every agent run afterwards, and it reads exactly as confident as a true one.”
Another commenter, Juraj Madzunkov, asked the sharper operational question: “Does anything actually retire one of those ‘we tried X’ facts, or does it just sit there until a human notices it’s wrong?” Asad’s conclusion was blunt: “Nothing retires it today as far as I can tell, and that’s the gap.”
Greplica’s maker did not fully answer that in the launch thread. The “anchored in committed code” mechanism gives you a natural expiry for facts about code that still exists. But negative facts are not anchored in anything that changes. Rabnoor Singh, who runs a memory folder for his own agent setup, put it well: “Nothing in the repo changes when it stops being true. It just sits there getting more confident with age.”
Where the math breaks: the “we tried X” problem in e-commerce
This is the same failure mode I see in cross-border teams. Somebody tests TikTok Spark Ads with one weak creative, declares “TikTok doesn’t work for us,” and six months later every new hire and every AI assistant is confidently repeating that while a competitor quietly wins with the same ad budget.
The math breaks because negative facts are the ones that prevent action, and prevention is invisible. No dashboard ever screams that you did not do the thing. A dated note tells you to doubt. A reproducible check tells you to delete. Greplica gets the first half right. The second half — making negative facts re-testable — is still unsolved, and it is the difference between a memory that compounds and a memory that slowly poisons every decision.
What I’d watch / test next
This week, I would do three things. First, if you have a GitHub repo with any meaningful history, clone the Greplica repo and run it against a real project for two weeks. Do not wait for your engineering team to ask. Second, build a decision log for your own brand, even in a spreadsheet. Every time you change a listing, a return policy, an ad target, or a supplier, write the decision, the evidence, and the date it should be reviewed. Third, wire that log into whatever AI assistant reads your chat or help-desk exports, and force it to cite the log before answering an operational question. The test is not whether the log looks neat. The test is whether, thirty days from now, a new hire or an AI agent can answer “why do we do it this way” without asking you. Watch Greplica’s roadmap for one feature: the ability to attach an expiry or a reproducible check to negative facts. That is the moment this pattern becomes safe enough for serious e-commerce operations to trust.
