LLM Context Hygiene

When you work with an AI agent in Rakenne, everything in your project workspace and conversation history can influence its responses. Context hygiene means keeping that input relevant and focused so the agent stays on task, gives better answers, and uses fewer tokens.

Why it matters

• Accuracy — Irrelevant files or old messages can steer the model toward outdated or wrong assumptions.
• Cost and speed — Smaller, focused context means faster replies and lower token usage.
• Reproducibility — A clean context makes it easier to get consistent results when you repeat a task.

Automatic context compaction

Long document-elaboration sessions can grow until the conversation history approaches the limits of the model’s context window. Rakenne’s agent uses automatic context compaction so that long sessions stay within bounds without you having to start over.

When compaction runs

Compaction is triggered automatically in two situations:

• Threshold — The conversation has grown enough that context size is nearing the model’s limit. The agent compacts before hitting the ceiling so the next turn can still be processed.
• Overflow — Context has already exceeded the safe size (for example after a large paste or a long reply). The agent compacts to bring context back into range.

What happens during compaction

The agent summarizes the earlier part of the conversation and keeps recent messages and important details. In effect, older turns are replaced by a concise summary so that the session can continue with a smaller, focused context. You don’t lose the session or the project; you keep working in the same thread with a shortened history.

What you see in the chat

When compaction completes successfully, you’ll see a short inline message in the conversation, for example: “To keep things running smoothly, I’ve summarized our earlier conversation. If I seem to have forgotten something specific, just remind me.” That’s your cue that the agent has trimmed context. If something important was lost in the summary, you can briefly remind the agent and continue.

Compaction is fully automatic. You can still start a new session whenever you want a clean slate instead of relying on compaction.

How to keep context clean in Rakenne

1. Start new sessions for new tasks
Use a fresh project or a new session when the goal changes. That way the agent isn’t carrying over assumptions from a previous, unrelated task.

2. Keep the workspace tidy
Move or remove files you don’t need for the current task. The agent can only use what’s in the project; fewer irrelevant files mean less noise.

3. Be explicit in your prompts
State what you want and what to use (e.g. “using only the data in report.csv”). That reduces the chance the model will pull in the wrong sources.

4. Use session management
Rakenne’s session model lets you control when context is carried over. Learn how sessions work and when to start a new one — see the Session management tutorial.

Good context hygiene is a habit: small, consistent choices that keep your AI interactions focused and efficient.