01 · Search
Search finds what matches the words.
- README: money handling, two paragraphs
- A forum thread: float or decimal for currency
- Issue: rounding bug in refunds, closed last year
For engineering teams
Your coding agent has your docs. Give it your judgment.
Your review comments say what you block and why. Duplikate turns them into rules your coding agents load before they work, scoped to each repository, and grades them by replaying your past reviews.
The gap
Docs say how. Reviews say what you will not accept.
A coding agent reads the README, the code and whatever a search returns. It does not know that money is integer cents in this codebase, that the refund flow has been burned before, or which shortcuts you always send back.
Your review comments know. So do the corrections you make to agent output, and the decisions in your pull request threads. That record is what an engineering engram is built from.
What becomes a rule
From your reviews to rules an agent follows.
Your review comments
What you asked for, what you blocked, and the reason you gave. The clearest record of your rules that exists.
Your corrections to agent output
Every edit you make to an agent's change says what it should have done. Each becomes a proposed rule you keep or discard.
Your decisions in threads
Why the queue is idempotent, why a migration runs in two steps. Recorded once, and cited every time it applies.
Scoped per repository
A rule from one codebase stays there unless you say otherwise. Names of people and companies become roles at the door.
See it learn
Same question, three layers.
A fictional person, Sam. Step through the layers. Nothing on this page reads anything of yours.
Sam leads a small engineering team.
The questionA pull request stores an order total as a float.
02 · Graph
A graph connects what relates to what.
- The pull request
- The file: order totals
- Called by the refund flow
- Test coverage on that path: partial
- The author: new to the team
03 · Sam's engram
Only Sam knows what to do with it. Sam's engram does too, and shows why.
Request changes: store money as integer cents.
Money is integer cents, always. Six past reviews said so.
Sam's review comments, six of them, over two years
The refund rounding bug of last year is why the rule exists.
Nice work on the flow. One blocker: totals must be integer cents, not floats; see the refund rounding bug from last year. Change the column and the two call sites and I'll approve.
Drafted by Sam's AI. Sam approves before it posts.
How it remembers
Stored once, recalled with the source.
Stored · 3 March
A review comment Sam wrote on a pull request
Totals are integer cents. A float here is how last year's refund rounding bug started.
- type
- rule
- claim
- Money is integer cents, always.
- source
- characters 0 to 25: "Totals are integer cents."
- valid from
- 3 March
Recalled · Twelve days later
A new question, in other words
A new pull request adds a discount field as a decimal.
Money is integer cents, always.Sam's review comment of 3 March
Request changes: keep the discount in integer cents, as the totals are.
Drafted by Sam's AI. Sam approves before it posts.
How it works
Graded by replaying your past reviews.
- 01Choose the repositoriesYour review comments, corrections and decisions, in the repositories you choose. Comments by others are read only as the context your reply answered.
- 02We extract your rulesEach comment that states a rule becomes a node, citing the review and the change it was about.
- 03We replay your reviewsPast pull requests are replayed with your comments held out: would the agent have blocked what you blocked? You see every miss.
- 04Your agents load the rulesThe rules for a repository load before an agent works there. Every block cites the review it came from, and you approve what it learns next.
Side by side
Three ways to give an agent context.
| Dimension | Repo docs and search | An agent memory layer | An engineering engram |
|---|---|---|---|
| What it knows | What was written down | What came up in its own sessions | What you decided in review, with your reason |
| Where a rule comes from | Whoever wrote the doc | Its own summaries | Your own comments, each cited |
| How it is checked | It is not | Vendor-run benchmarks | Replayed against your past reviews |
| How it stays current | When someone updates the doc | As sessions pile up | From every review and correction you make |
Categories, not products: what each kind of tool does by design, and how agent-memory vendors reported results on their public pages, read 24 October 2026.
Who it is not for
When the answer is not this.
Your team does not review code. There is nothing to learn from without reviews.
You want rules from a style guide. This learns from what you actually block, not from what a guide says.
You want a team's rules learned from people who did not agree to it. An engineering engram is built from its owner's own comments only.
Questions
What people ask first.
- Does it train a model on our code?
- We train no model on you or your code. Your rules are nodes you can read, cite and delete, retrieved when an agent works. Your record is never used to train general AI models, ours or anyone else's.
- Which coding agents does it work with?
- Any agent that can load project rules or instructions before it works. We set it up with your team as part of the build, and name no vendor on this site.
- Whose comments does it learn from?
- The owner's own. Comments by other people are read only as the context a reply answered, and never become rules. A colleague who wants an engram builds one from their own record.
- Does it copy our code?
- A rule holds the reasoning, not the code. Each node cites its source by an identifier and a span, rather than storing what it came from.
- How do we know the rules are right?
- We replay past pull requests with your comments held out, and count how often the agent would have blocked what you blocked. You see every miss before anything runs.
Next
Read the engineering, or start.
Talk to us
Talk to an engineer.
You'll talk to a senior engineer, not a sales rep, and leave knowing what your engram would be built from and how it would be measured.