Context Compactor

Automatic context compaction for OpenClaw when using local models that don't properly report token limits or context overflow errors.

The Problem

Cloud APIs (Anthropic, OpenAI) report context overflow errors, allowing OpenClaw's built-in compaction to trigger. Local models (MLX, llama.cpp, Ollama) often:

• - Silently truncate context
• Return garbage when context is exceeded
• Don't report accurate token counts

This leaves you with broken conversations when context gets too long.

The Solution

Context Compactor estimates tokens client-side and proactively summarizes older messages before hitting the model's limit.

How It Works

CODEBLOCK0

Installation

CODEBLOCK1

The setup command automatically:

• - Copies plugin files to INLINECODE0
• Adds plugin config to openclaw.json with sensible defaults

Configuration

Add to openclaw.json:

CODEBLOCK2

Options

Option	Default	Description
INLINECODE3	INLINECODE4	Enable/disable the plugin
INLINECODE5

8000 | Max context tokens before compaction | | keepRecentTokens | 2000 | Tokens to preserve from recent messages | | summaryMaxTokens | 1000 | Max tokens for the summary | | charsPerToken | 4 | Token estimation ratio | | summaryModel | (session model) | Model to use for summarization |

Tuning for Your Model

MLX (8K context models):
CODEBLOCK3

Larger context (32K models):
CODEBLOCK4

Small context (4K models):
CODEBLOCK5

Commands

/compact-now

Force clear the summary cache and trigger fresh compaction on next message.

CODEBLOCK6

/context-stats

Show current context token usage and whether compaction would trigger.

CODEBLOCK7

Output:
CODEBLOCK8

How Summarization Works

When compaction triggers:

1. 1. Split messages into "old" (to summarize) and "recent" (to keep)
2. Generate summary using the session model (or configured summaryModel)
3. Cache the summary to avoid regenerating for the same content
4. Inject context with the summary prepended

If the LLM runtime isn't available (e.g., during startup), a fallback truncation-based summary is used.

Differences from Built-in Compaction

Feature	Built-in	Context Compactor
Trigger	Model reports overflow	Token estimate threshold
Works with local models

❌ (need overflow error) | ✅ | | Persists to transcript | ✅ | ❌ (session-only) | | Summarization | Pi runtime | Plugin LLM call |

Context Compactor is complementary — it catches cases before they hit the model's hard limit.

Troubleshooting

Summary quality is poor:

• - Try a better INLINECODE17
• Increase INLINECODE18
• The fallback truncation is used if LLM runtime isn't available

Compaction triggers too often:

• - Increase INLINECODE19
• Decrease keepRecentTokens (keeps less, summarizes earlier)

Not compacting when expected:

• - Check /context-stats to see current usage
• Verify enabled: true in config
• Check logs for [context-compactor] messages

Characters per token wrong:

• - Default of 4 works for English
• Try 3 for CJK languages
• Try 5 for highly technical content

Logs

Enable debug logging:

CODEBLOCK9

Look for:

• - INLINECODE24
• INLINECODE25

Links

• - GitHub: https://github.com/E-x-O-Entertainment-Studios-Inc/openclaw-context-compactor
• OpenClaw Docs: https://docs.openclaw.ai/concepts/compaction