MEMORY.md is fully injected into system prompt despite memory_search/memory_get being available

Summary

MEMORY.md is unconditionally injected in full into the system prompt as part of Project Context → Workspace Files. At the same time, the system prompt also instructs the model to use memory_search and memory_get tools for memory recall. This creates a paradox: if the model is expected to search memory on demand, why is the entire memory file pre-loaded into the context on every request?

For users who accumulate substantial long-term memory (rules, milestones, project states, lessons learned), MEMORY.md can easily grow to 9,000+ characters (~2,300–3,800 tokens), making it one of the largest contributors to system prompt cost.

Observed behavior

In a captured request payload from a clean conversation:

• MEMORY.md full content (~9k chars) is injected under # Project Context in the system prompt
• The system prompt also contains ## Memory Recall rules instructing the model to run memory_search → memory_get before answering memory-related questions
• memory_search and memory_get are both available in body.tools[]

This means the model receives the same information twice: once passively via system prompt injection, and once actively via the search/retrieval tools.

Expected behavior

Ideally, users should have the option to control whether MEMORY.md is fully injected or only accessible via memory_search/memory_get. A possible approach:

Option A: Configuration flag

Add a workspace-level or agent-level config option, e.g.:

workspace_files:
  MEMORY.md:
    inject: false  # do not inject into system prompt; rely on memory_search/memory_get

When inject: false, the file is still indexed by memory_search but not included in the system prompt.

Option B: Encourage content tiering (documentation)

As an interim solution, document the best practice of keeping MEMORY.md lean (only critical, high-frequency rules) and moving detailed/historical content to memory/*.md daily log files, which are not auto-injected but are searchable via memory_search.

Impact

For power users on pay-per-token providers, this adds up quickly across hundreds of daily interactions.

Environment

• OpenClaw: latest (self-hosted)
• Debug method: custom model provider proxy (modelbox) to capture full request payloads