Every LLM has a context window limit. When your agent exceeds it, responses degrade — hallucinations increase, instructions get forgotten, and costs skyrocket. This guide teaches you to manage context like a pro: what to keep, what to compress, and what to offload.

Why Context Matters

Cost scales with tokens

Every token in your context window costs money on every API call. A bloated 100K context at $15/M tokens = $1.50 per request.

Quality degrades with noise

LLMs perform worse with irrelevant context. Studies show accuracy drops 20-30% when context is cluttered with irrelevant information.

Speed decreases linearly

Larger context = slower responses. Time-to-first-token increases proportionally with context size.

Anatomy of OpenClaw Context

Component	Tokens	Priority	Notes
SOUL.md	~500-2K	Always loaded	Agent identity and core instructions. Keep concise.
AGENTS.md	~300-1K	Always loaded	Role definitions. Grows with multi-agent setups.
memory.md	~1K-10K	Always loaded	Biggest culprit for bloat. Must be actively managed.
Active conversation	~2K-50K	Dynamic	Current chat history. Auto-managed by sliding window.
Tool results	~1K-20K	Per-call	Web scrapes, file reads, search results. Can explode quickly.

Optimization Strategies

1. Prune memory.md weekly

Review memory.md every week. Remove outdated facts, merge duplicates, and move structured data to SQLite. Target: under 2K tokens.

openclaw memory status  # Check current token count

2. Use tiered context loading

Not everything needs to be in context always. Use @include directives to load sections conditionally based on the current task.

# In SOUL.md:
@include(coding-rules.md) when task.type == 'code'
@include(writing-style.md) when task.type == 'write'

3. Summarize tool outputs

Web scrapes can be 10K+ tokens. Use the built-in summarizer to compress results before injecting into context.

openclaw config set tools.web.summarize true
openclaw config set tools.web.max_tokens 2000

4. Set conversation sliding window

Limit how many previous messages stay in context. Older messages get summarized automatically.

openclaw config set context.sliding_window 20
openclaw config set context.summary_after 10

5. Offload to external memory

Use SQLite or vector databases for large knowledge bases. Query on-demand instead of keeping everything in context.

openclaw memory index --to sqlite
# Queries memory only when relevant

Token Budget Template

Component	Budget	Tip
SOUL.md	800	One page max. Every word must earn its place.
AGENTS.md	500	Brief role descriptions. Details go in per-agent files.
memory.md	1,500	Top 30 facts only. Structured data → SQLite.
Conversation history	8,000	20-message sliding window with auto-summarization.
Tool results	3,000	Summarize all web/file results. Cap per-tool output.
Reserved for response	4,000	Always leave room for the model to think and respond.

Anti-Patterns to Avoid

Dumping entire files into memory.md

Store file paths. Let the agent read files on-demand.

Keeping full conversation history forever

Use sliding window. Summarize older turns.

Raw HTML from web scrapes in context

Always extract text and summarize before injecting.

Duplicate information across SOUL/AGENTS/memory

Single source of truth. Reference, don't copy.

Verbose system prompts with examples

Keep prompts concise. Move examples to on-demand retrieval.

Monitoring Context Usage

openclaw config get agents.defaults # Current context breakdown

openclaw memory status # Token count per section

openclaw memory status # memory.md token count

openclaw config get context.* # All context settings

Optimization Checklist

✅ memory.md under 2K tokens

✅ SOUL.md under 1K tokens

✅ Sliding window set (20 messages)

✅ Tool output summarization enabled

✅ Structured data migrated to SQLite

✅ Conditional @include for context sections

✅ Response token budget reserved (4K+)

✅ Weekly memory.md review scheduled

Why Context Matters

Cost scales with tokens

Every token in your context window costs money on every API call. A bloated 100K context at $15/M tokens = $1.50 per request.

Quality degrades with noise

LLMs perform worse with irrelevant context. Studies show accuracy drops 20-30% when context is cluttered with irrelevant information.

Speed decreases linearly

Larger context = slower responses. Time-to-first-token increases proportionally with context size.

Anatomy of OpenClaw Context

Component	Tokens	Priority	Notes
SOUL.md	~500-2K	Always loaded	Agent identity and core instructions. Keep concise.
AGENTS.md	~300-1K	Always loaded	Role definitions. Grows with multi-agent setups.
memory.md	~1K-10K	Always loaded	Biggest culprit for bloat. Must be actively managed.
Active conversation	~2K-50K	Dynamic	Current chat history. Auto-managed by sliding window.
Tool results	~1K-20K	Per-call	Web scrapes, file reads, search results. Can explode quickly.

Optimization Strategies

1. Prune memory.md weekly

Review memory.md every week. Remove outdated facts, merge duplicates, and move structured data to SQLite. Target: under 2K tokens.

openclaw memory status  # Check current token count

2. Use tiered context loading

Not everything needs to be in context always. Use @include directives to load sections conditionally based on the current task.

# In SOUL.md:
@include(coding-rules.md) when task.type == 'code'
@include(writing-style.md) when task.type == 'write'

3. Summarize tool outputs

Web scrapes can be 10K+ tokens. Use the built-in summarizer to compress results before injecting into context.

openclaw config set tools.web.summarize true
openclaw config set tools.web.max_tokens 2000

4. Set conversation sliding window

Limit how many previous messages stay in context. Older messages get summarized automatically.

openclaw config set context.sliding_window 20
openclaw config set context.summary_after 10

5. Offload to external memory

Use SQLite or vector databases for large knowledge bases. Query on-demand instead of keeping everything in context.

openclaw memory index --to sqlite
# Queries memory only when relevant

Token Budget Template

Component	Budget	Tip
SOUL.md	800	One page max. Every word must earn its place.
AGENTS.md	500	Brief role descriptions. Details go in per-agent files.
memory.md	1,500	Top 30 facts only. Structured data → SQLite.
Conversation history	8,000	20-message sliding window with auto-summarization.
Tool results	3,000	Summarize all web/file results. Cap per-tool output.
Reserved for response	4,000	Always leave room for the model to think and respond.

Anti-Patterns to Avoid

Dumping entire files into memory.md

Store file paths. Let the agent read files on-demand.

Keeping full conversation history forever

Use sliding window. Summarize older turns.

Raw HTML from web scrapes in context

Always extract text and summarize before injecting.

Duplicate information across SOUL/AGENTS/memory

Single source of truth. Reference, don't copy.

Verbose system prompts with examples

Keep prompts concise. Move examples to on-demand retrieval.

Context Window Optimization

Why Context Matters

Cost scales with tokens

Quality degrades with noise

Speed decreases linearly

Anatomy of OpenClaw Context

Optimization Strategies

1. Prune memory.md weekly

2. Use tiered context loading

3. Summarize tool outputs

4. Set conversation sliding window

5. Offload to external memory

Token Budget Template

Anti-Patterns to Avoid

Monitoring Context Usage

Optimization Checklist

▶ Related Resources

Context Window Optimization

Why Context Matters

Cost scales with tokens

Quality degrades with noise

Speed decreases linearly

Anatomy of OpenClaw Context

Optimization Strategies

1. Prune memory.md weekly

2. Use tiered context loading

3. Summarize tool outputs

4. Set conversation sliding window

5. Offload to external memory

Token Budget Template

Anti-Patterns to Avoid

Monitoring Context Usage

Optimization Checklist

▶ Related Resources