Answer
How do you preserve important decisions between AI sessions?
Write each one down with the reason behind it, append rather than edit, and mark the ones that get reversed.
Record each decision with the reason it was taken, in a file that is appended to rather than rewritten. The reason is what lets a later reader tell whether the decision still holds, which the decision alone cannot.
The failure this solves is specific and expensive: a choice is made carefully in one session, and in a later session the same question is reopened and answered differently, with no awareness that it was ever settled. The work then contains two incompatible assumptions, and the discovery usually comes from something breaking rather than from anyone noticing the contradiction.
The record has to contain the reason, not just the choice. A line saying the system uses one identifier rather than another is a rule. The same line saying it uses that identifier because the other is not unique across regions is a rule with a test attached: a later reader can check whether that is still true and decide accordingly. Without the reason, a later session either treats the decision as immovable when circumstances have changed, or overturns it because the original constraint was invisible.
The file should be appended to rather than edited. Rewriting produces a document that always describes the current position and never shows the history, which sounds tidier and removes the information that matters most: that something was decided, reconsidered and changed. Where a decision is superseded, the honest form is a new entry that names the one it replaces and says why, leaving both in place.
Dates matter more here than in other documents. A decision from three months ago and one from last week carry different weight when the surrounding situation has moved, and a reader cannot apply that judgement without knowing which is which. This is also what makes periodic review possible: an entry that is a year old and has never been checked is a specific, findable thing rather than a vague worry.
Keep it short and keep it about decisions. The most common way these records fail is by becoming a general project journal, at which point nobody reads them and the decisions are buried in narrative. If a line does not record a choice between alternatives, it belongs elsewhere. A record of thirty decisions is readable in a few minutes and will be read; one of three hundred entries will be searched, and searching only finds what you already suspected.
One thing to be explicit about: this is a record for humans and systems alike, and it should be supplied at the start of sessions that could reopen a settled question. A file nobody reads at the right moment is a record of having been careful rather than an instrument that prevents the failure.
A decision recorded without its reason is a rule with no expiry test, and it will be either obeyed too long or overturned too easily.
Siddharth Sharma, Context Theory
Related questions
Should the AI maintain this file itself?
It can draft entries and the entries should be read before they land. The specific hazard is that a system records its interpretation of a decision rather than the decision, and the interpretation is what gets applied indefinitely afterwards. Reading the addition takes seconds and is the difference between a record and a game of telephone.
What counts as important enough to record?
Anything you would be annoyed to have to decide again, and anything where a later reversal would create inconsistency rather than merely a different result. That is a narrower set than it sounds — most choices in a working day are local and leave no trace worth preserving — and the ones that qualify are usually obvious in the moment they are made.
METHOD
Every figure below carries its source and the date it was verified. Nothing on this page is asserted.
The numbers on this page.
| What | Value | Specific to |
|---|---|---|
| Close rate — response under 5 minutes vs over 24 hours | 32% vs 12% | Category-wide |
| Sub-15-minute compliance — automated routing vs manual only | 62.5% vs 39.1% | Category-wide |
Optifai speed-to-lead benchmark · n=939 companies · Q2 2025–Q1 2026 · verified
2026 speed-to-lead benchmark · verified
What is specific to this page.
| Kind | Claim | Check it against |
|---|---|---|
| Workflow | A decision recorded without its reason gives a later reader no test of whether it still applies, producing either continued adherence after the constraint has lapsed or reversal because the original constraint was invisible. | Taking any recorded decision without a stated reason and attempting to determine whether it still applies. |
| Response | An append-only record with explicit supersession preserves the fact that something was decided, reconsidered and changed, which a rewritten current-state document removes while appearing tidier. | Comparing an edited decisions document against its version history for reversals that are no longer visible in the text. |
| Constraint | Dating entries is what makes periodic review possible, because an unchecked entry of a given age becomes a findable object rather than a general concern about staleness. | Filtering a dated decisions record for entries older than a chosen interval that have not been revisited. |
| Software | Decision records fail by becoming general project journals, after which they are searched rather than read, and searching surfaces only what the reader already suspected while reading surfaces contradictions. | Counting entries in an existing record that do not document a choice between alternatives. |
Each row would be wrong on another industry's page. Where a sourced figure exists it is in the table above instead; these are the constraints that shape the work and do not happen to be numbers.
Start with the measurement.
Reading about a benchmark is not the same as knowing your own number. The audit produces yours, measured rather than estimated.
$497 · delivered in 5 business days · credited against month one