Answer
What should an agent write down and what should it look up?
Write down what leaves no trace elsewhere. Look up anything with an authoritative source, because a copy goes stale silently.
Write down what exists nowhere else: decisions, the reasons behind them, and what was tried and failed. Look up anything with an authoritative source, because a written copy goes stale silently and the source does not.
The division is about where truth lives. Some things have an authoritative home — the record, the file, the system of record, the current state of the work — and copying them into a note creates a duplicate that begins diverging immediately. Other things have no home at all: why an approach was chosen, what was ruled out, what an ambiguity was resolved as. If those are not written, they are gone at the end of the session and will be re-derived, differently.
So the test is traceability. Could this be established by looking at something? If yes, look at it, every time, and hold a pointer rather than a copy. If no, write it down, because there is nothing to look at. This single question resolves most of the cases people find difficult, and it explains why so many project notes are worse than useless: they are copies of things that changed.
Three categories reliably belong in writing. Decisions with their reasons, which are the most expensive thing to reconstruct and the most damaging to lose. Negative findings — what was tried, what failed and why — which are pure cost to rediscover and are almost never recorded. And interpretations of ambiguity: where the brief could have meant two things and one was chosen, the choice is a fact about the work that exists nowhere else.
Three categories reliably belong in lookups. Current state of any system, which is definitionally out of date the moment it is copied. Content of documents that have an owner, since the owner will change them without telling you. And anything computed from data, because the data moves and the computation is cheap to repeat.
The hard case is a summary of something large: a report on a codebase, an analysis of a dataset, a synthesis of several documents. It is derived, so by the rule it should be recomputed, and it was expensive to produce, so people keep it. The workable compromise is to keep it with a date and a note of what it was derived from, treat it as orientation rather than as fact, and re-derive it before any decision that turns on a specific inside it.
One practical note on where things get written. A note that lives in the session is not written down; it is present until it is not. Writing means a file, in a location the next session reads, with a name that says what it is. The distinction sounds pedantic and it is the difference between the practice working and appearing to.
Copy a fact and you have created a second version of it that will be wrong before anyone notices.
Siddharth Sharma, Context Theory
Related questions
Should an agent keep notes for its own use during a run?
Yes, and it is one of the more effective things it can do on long work. Notes written to a file during a task survive compaction and session boundaries, and re-reading a note is more reliable than recalling a summary of a summary. The same traceability rule applies: notes about what was found are valuable, notes copying what a file contains are not.
How do you stop the written record becoming its own maintenance problem?
By keeping it to the three categories, which are naturally self-limiting: most work involves few genuine decisions and few dead ends. Records grow unmanageable when they become journals of activity, and the corrective is the same test — if it could be established by looking at the work, it does not belong in the record.
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 |
| Visibility lift in AI-generated answers from GEO methods | up to 40% | Category-wide |
Optifai speed-to-lead benchmark · n=939 companies · Q2 2025–Q1 2026 · verified
Aggarwal et al., "GEO: Generative Engine Optimization", Princeton / Georgia Tech / IIT Delhi / Allen Institute for AI — KDD 2024 · GEO-bench · 10,000 queries across 8 domains · verified
What is specific to this page.
| Kind | Claim | Check it against |
|---|---|---|
| Workflow | Copying a fact that has an authoritative source creates a second version that begins diverging immediately and reports nothing when it does, which is why so many project notes are actively misleading rather than merely redundant. | Comparing each copied statement in an existing project note against its source of record. |
| Response | Decisions with reasons, negative findings and resolutions of ambiguity exist nowhere but in the working session, so their loss is total at session end and they are re-derived differently rather than recovered. | Attempting to reconstruct why a past choice was made from the artefacts alone. |
| Software | Derived summaries of large material occupy a middle position: keeping them with a derivation date and treating them as orientation rather than fact preserves their value while requiring re-derivation before any decision turning on a specific within them. | Checking a retained summary against the material it was derived from for specifics that have since changed. |
| Constraint | A note held in the session is not a written record, because writing requires a file in a location the next session reads under a name that identifies it, and the difference determines whether the practice works or only appears to. | Starting a fresh session and checking whether the note is available without a person supplying it. |
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