Context Engineering Principles

Context Engineering Principles

The principles behind file-based planning, distilled from Manus's context engineering approach.

Filesystem as External Memory

Context Window = RAM (volatile, limited)
Filesystem     = Disk (persistent, unlimited)

Anything important gets written to disk. The context window is temporary -- it resets, it fills up, it loses older content. Files persist across sessions, have no size limit, and can be selectively re-read when needed.

When compressing context (summarizing, dropping details), always preserve pointers to the full data: keep URLs even if web content is dropped, keep file paths when dropping document contents. Never lose the ability to recover the original.

Manipulate Attention Through Recitation

This is the key insight behind the entire approach.

After ~50 tool calls, models forget original goals. This is the "lost in the middle" effect -- content at the start and end of the context window gets attention, but the middle fades.

The fix: re-read task_plan.md before each major decision. This pushes the goal and current state into the most recent part of the context, where it gets maximum attention.

Start of context: [Original goal -- far away, low attention]
...many tool calls in the middle...
End of context: [Recently read task_plan.md -- HIGH attention]

This is why Manus can handle ~50 tool calls without losing track. The plan file acts as a goal-refresh mechanism.

Keep the Wrong Turns In

Leave failed attempts and error traces in the planning files.

Why:

• Failed actions with stack traces let the model implicitly update its beliefs about what works
• Reduces repetition of the same mistakes
• Error recovery is one of the clearest signals of effective agentic behavior

The error table in task_plan.md serves this purpose. Log every failure with the attempt number and what you tried. The next attempt should always be different.

Avoid Repetitive Patterns

Repetitive action-observation pairs cause drift and hallucination. When you notice yourself doing the same thing repeatedly:

• Vary your approach (different tool, different angle)
• Re-read the plan to check if you're still on track
• If stuck after 3 attempts, escalate to the user

Context Offloading

Store full results on disk, keep only references in context.

Practical application:

• Write large search results or API responses to files
• Keep file paths and summaries in context
• Use glob and grep to find information later
• Load details only when needed (progressive disclosure)

This keeps the context window focused on the current task while making all accumulated knowledge accessible through the filesystem.

Source

Based on context engineering principles from Manus.