// Section 8 — Autocompact · 3 MIN READ

[✓] VERIFIED MANUAL ENTRY — This concept has been rewritten from primary sources and is legally cleared for production.

Autocompact

Compaction triggered automatically by the client harness when context size crosses a threshold (often 80%), risking the quiet loss of task constraints.

Autocompact is an automated performance feature built into some agent harnesses. When the accumulated tokens in the context window cross a specific threshold (typically 80% capacity), the harness automatically runs a summarization query, clears the history, and injects the summary back as the new starting state — without asking the user.

You interface with autocompact when:

  • Running long-running CLI agents that continue executing tasks in a loop.
  • Building chatbot platforms that must manage memory without crashing.

The Autocompact Risk: Quiet Memory Loss

While autocompact prevents token overflow errors, it poses a significant development risk: it fires at a moment you did not choose.

  • The Lossy Summary: If the threshold is crossed mid-task (e.g. halfway through a complex code refactor), the model writes its own summary. It decides which constraints to preserve, occasionally discarding crucial details (like a type requirement or user instruction).
  • Silent Failures: The agent carries on working with complete confidence, but has quietly forgotten a constraint you established earlier in the session. You only notice the failure when its code starts violating your specifications.
Session Hits 80% Token Limit ──► Autocompact Fires (Automatic Summary) ──► Crucial Constraint Lost ──► Silent Bug

Defending Against Autocompact

To prevent autocompact from losing your decisions:

  1. Monitor Context Indicators: Watch the token size indicator in your console and compact manually at a natural milestone (e.g. between tasks).
  2. Write to Disk: Keep project requirements and decisions written inside a file on your disk (like a plan.md spec). No automated summary can wipe files stored on your filesystem.
  3. Adjust Thresholds: Configure your harness settings to increase the buffer or turn autocompact off entirely.

# AVOID

Do not rely on conversation logs to store your project's styling and API constraints during a long multi-turn session. Autocompact will eventually wipe the detailed logs, diluting the rules.

  • Avoid: Pasting your system rules once and assuming the agent will remember them for 100 turns.
  • Write: Keep global rules in AGENTS.md or a local file, where they are loaded fresh on every turn regardless of compaction.

# USAGE

Developer A: "The model suddenly started ignoring our Postgres naming conventions that we discussed in the first message." Developer B: "Check the logs. Autocompact fired automatically between turns, and the generated summary lost track of those naming rules. Let's write the rules into AGENTS.md so they stay locked in context."

// SEE_ALSO

CompactionAn in-memory session reset where the active chat history is summarized by the model, throwing away the detailed transcript to free up context window space.
Context WindowThe absolute maximum token limit (capacity) that a model can process, read, and write in a single API request.
Attention DegradationThe gradual decline in a model's constraint-following and reasoning performance as prompt context length increases, caused by attention budget dilution.
Handoff ArtifactA persistent file written to the environment by an agent to record plans, status, and decisions, used to brief a fresh successor session.

// SOCRATIC_VALIDATION

Interactive Concept Quiz

QUESTION 1 OF 3SCORE: 0/3

[EDIT_THIS_TERM_ON_GITHUB ↗]