Answer
How should context be compressed without losing important information?
Decide what must survive before compressing, and keep those items verbatim. A summary chooses what to lose, and nobody is consulted.
Name what must survive before you compress, and carry those items forward verbatim rather than summarised. Summarisation removes specifics silently, and the specifics are what later decisions depend on.
The trouble with compression is not that it loses information; it is that the losses are unannounced and unchosen. A summary reads as complete because it is coherent, and the values, exceptions, names and edge cases it dropped leave no gap where they were. Later work then proceeds on a foundation that appears whole, and the missing specific surfaces as a wrong answer rather than as a question.
The workable discipline is to decide the survivors first. Before any compression, name the categories that must come through intact: identifiers, values that other work depends on, decisions and their reasons, constraints, and anything discovered that was expensive to discover. Those are carried verbatim. Everything else may be summarised, and the loss is then bounded by a choice rather than by a process.
The second principle is that a summary should preserve specificity over completeness. A short account containing three exact facts is more useful than a longer one that covers everything approximately, because approximate coverage is exactly what the system can regenerate on its own and exact facts are what it cannot. Summaries that read as balanced overviews are usually the least useful kind.
A structural alternative is better than compression wherever it is available: write to durable storage as you go and re-read rather than compress. A finding written to a file at the moment it was established survives at full fidelity, and the session can hold a pointer instead of the content. This turns a lossy operation into a lookup, and it is why note-taking during long work is a reliability practice rather than an administrative one.
Where automatic compaction is applied by a tool, the practical question is what it did. Reading the summary it produced, once, on a job that matters, is worth the two minutes: it tells you what class of thing this mechanism discards, and that generalises. Teams that have never read one are relying on a transformation they have not inspected, which is the same position as trusting an output because it sounded right.
Finally, a compressed context should record that it is compressed and what it stands for. A note saying this summarises fourteen exchanges about the migration, full detail in a named file, costs one line and changes how the next reader treats it. Without it, a summary is indistinguishable from a complete account, and it will be relied on as one.
Every summary is a decision about what to forget, taken by something that will not tell you what it chose.
Siddharth Sharma, Context Theory
Related questions
Is it better to compress or to start a new session?
Starting fresh from written state is better whenever the state was written, because the durable record was chosen deliberately while a summary was not. Compaction is the right tool when the work is genuinely mid-flight and nothing has been written down, which is itself a signal about how the work was structured.
What gets lost most often?
Exact values, the reasoning behind choices, and negative findings — the things that were tried and did not work. The last is the most costly because it is invisible: nothing in the summary indicates that a route was already explored, so it gets explored again at full price by whatever continues the work.
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 | Compression losses are unannounced and unchosen, so a summary reads as complete because it is coherent while the values, exceptions and edge cases removed leave no visible gap, and the omission surfaces later as a wrong answer rather than as a question. | Comparing a generated summary against the material it replaced for specifics that did not survive. |
| Response | Naming the categories that must survive before compressing — identifiers, depended-upon values, decisions with reasons, constraints and expensive discoveries — bounds the loss by a choice rather than by a process. | Listing required survivors before compaction and checking each against the resulting summary. |
| Software | Writing findings to durable storage during the work converts a lossy compression into a lookup, because a fact recorded at the moment of discovery survives at full fidelity while the session holds only a pointer. | Comparing detail retained through compaction against detail retrieved from a file written during the same work. |
| Constraint | Negative findings are the most costly compression loss because nothing in a summary indicates that a route was already explored, so the exploration is repeated at full cost by whatever continues the work. | Checking whether a compacted session's summary records any approach that was tried and abandoned. |
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