Summary

The instructions in the hardcoded system prompt do not match the instructions in AGENTS.md.

Expected behavior

The prompts should not conflict with each other.

Actual behavior

The system prompt asks for selective memory reading via tool calls, while AGENTS.md instructs the agent to read entire files.

OpenClaw version

2026.2.26

Logs, screenshots, and evidence

OpenClaw's hardcoded system prompt in src/agents/system-prompt.ts says to use tools to get only needed lines from memory:

Before answering anything about prior work, decisions, dates, people, preferences, or todos: run memory_search on MEMORY.md + memory/*.md; then use memory_get to pull only the needed lines.

AGENTS.md says to just read the entire files directly:

Before doing anything else:
  1. Read SOUL.md
  2. Read USER.md
  3. Read memory/YYYY-MM-DD.md (today + yesterday)
  4. If in MAIN SESSION: Also read MEMORY.md

Impact and severity

The agent has to decide whether to follow the hardcoded system prompt instructions that return less context or the instructions in AGENTS.md which return more context. It's not clear which will win. The fact that AGENTS.md has these instructions at all though makes users assume they're not being contradicted elsewhere.