Direct answer: Compress Leela context with a two-level record rather than deleting arbitrary old moves. Keep the original intention, rules, and turning points verbatim; reduce routine early moves to one structured line; preserve the latest two or three moves in greater detail. Remove duplicate cell descriptions, technical identifiers, and unnecessary personal data. After compression, verify that a reader can reconstruct the sequence of decisions and distinguish the player’s words from AI hypotheses.
Shorter is not automatically more faithful
A long session creates two kinds of load: the model receives repeated material, and the player has to check whether an important distinction disappeared. Cutting off the beginning solves the size problem but may destroy causality. A late answer can look independent even though it emerged after a particular transition, objection, or experiment.
Reliable compression preserves the decision structure, not every sentence.
A five-field record for an ordinary move
A routine move can fit on one line:
Move 7 · cell 31 “Curiosity” · no transition · answer: “I avoid asking for the criteria” · action: request criteria by Friday
If the guide’s question changed the response, add it as a sixth field. If the player rejected an interpretation, record that too: interpretation rejected — it does not match the facts.
Do not merge the player’s words with the system’s paraphrase. Quotation marks and labels such as answer, hypothesis, and action preserve provenance.
Information that should survive compression
- the original intention and any formal change to it;
- the rules of the particular board version;
- arrows, snakes, and other route-changing transitions;
- direct wording referred to later in the session;
- contradictions and rejected prompts;
- chosen actions, deadlines, and resulting feedback;
- the open question that still has no answer.
Repeated cell descriptions, internal IDs, full system instructions, and interface duplicates rarely add reflective value.
A two-layer history
Layer one: persistent summary
This contains the intention, boundaries, key turning points, and current working hypotheses. It changes slowly and only after the player reviews it.
Layer two: detailed recent window
The latest moves remain close to their original form because they shape the current question. As they age, routine detail is collapsed, while decisive quotes and events remain retrievable.
Before and after
Before: eight paragraphs repeating the cell text, three similar AI questions, two drafts of the answer, and technical request metadata.
After:
Intention: understand why I do not start the project.
Move 3 · “Control” · player linked the theme to wanting to eliminate criticism in advance.
Move 4 · “Resource” · named two free evenings; action: a 30-minute draft.
Feedback: draft completed, not submitted.
Open question: what is the player waiting for before publication?
The compact version does not claim that control is the true cause. It shows where the hypothesis came from and what happened after the action.
A reconstruction test
Read only the compressed context and answer four questions:
- Is the starting intention clear?
- Can you see why the direction of reflection changed?
- Are fact, quote, and interpretation distinguishable?
- Is the unfinished action or question visible?
If any answer is no, restore the missing fragment. Compression must not increase certainty: labels such as possible, the player suggested, and not verified may be more important than saving a line.
Privacy is part of optimization
Context reduction is also an opportunity to remove unnecessary personal data. Full names, addresses, document numbers, and another person’s private messages are rarely required to understand a move. Keep only what is necessary for the chosen task, and manage the retention of the full log separately.
A compact context is useful when it remains an auditable map of the game, not a polished narrative that can no longer be checked against the original notes.
Questions about this topic
Can I simply delete the first moves?
Usually not. An early move may explain a later change. Keep it as a short event and mark earlier turning points separately.
What should remain verbatim?
The original intention, direct quotes used by later reasoning, explicit disagreement, and decisions. Other material may be carefully paraphrased.
How many recent moves should stay detailed?
There is no universal number. Two or three recent moves plus any older event still active in the current question is a practical starting point.
Can AI produce the compressed summary itself?
Yes, but a person should compare it with the log. A model can overemphasize one detail, merge separate ideas, or turn a hypothesis into a fact.
Do technical session fields belong in the prompt?
Usually not. Internal IDs, precise timestamps, and raw service responses should be included only when a specific function requires them.
Related Leela guides
- How AI uses the full history of moves and answers in Leela
- How AI analyses a player’s answers in Leela
- How an AI guide creates the final Leela game summary
- Leela move history
- How progress is saved in online Leela
- How to resume an unfinished Leela game
- AI guide in Leela: benefits and limits
- Leela with an AI guide: how the full history of moves is used
FAQ
Can I simply delete the first moves?
Usually not. An early move may explain a later change. Keep it as a short event and mark earlier turning points separately.
What should remain verbatim?
The original intention, direct quotes used by later reasoning, explicit disagreement, and decisions. Other material may be carefully paraphrased.
How many recent moves should stay detailed?
There is no universal number. Two or three recent moves plus any older event still active in the current question is a practical starting point.
Can AI produce the compressed summary itself?
Yes, but a person should compare it with the log. A model can overemphasize one detail, merge separate ideas, or turn a hypothesis into a fact.
Do technical session fields belong in the prompt?
Usually not. Internal IDs, precise timestamps, and raw service responses should be included only when a specific function requires them.
Sources and editorial method
This material is intended for learning and self-reflection. It does not promise to predict the future and does not replace professional help.