openclaw - 💡(How to fix) Fix Feature request: mandatory pre-action memory retrieval hooks [1 participants]

Title: Feature request: mandatory pre-action memory retrieval hooks

Problem

Current memory architecture (supermemory, LCM, file-based notes) stores information reliably but retrieval is entirely model-initiated. The model decides when to check memory, and under task pressure it consistently skips retrieval — leading to repeated mistakes on things it has already documented.

This is the #1 frustration reported after 67 days of heavy daily use across trading automation, flight bookings, and system administration.

What happens today

1. Agent learns a lesson (e.g., "BRENTOIL is the correct key, not OIL")
2. Agent writes it to memory (MEMORY.md, supermemory, daily notes)
3. Next time the same situation arises, agent barrels forward without checking
4. Same mistake repeats

The storage layer works. The retrieval trigger doesn't exist.

Proposed solution: mandatory retrieval hooks

Add a configurable pre-action retrieval step in the agent loop — not optional, not model-initiated:

memory:
  preActionRetrieval:
    enabled: true
    triggers:
      - before: exec        # before any shell command
        query: "relevant rules for {command_context}"
      - before: trade        # before trade-related actions
        query: "trading rules and recent lessons"
      - before: format       # before formatting output
        query: "output format preferences"
    sources:
      - supermemory
      - lcm
      - files: ["MEMORY.md"]
    maxTokens: 500

The key insight: retrieval should be architectural, not behavioral. Don't trust the model to remember to check — force it.

Alternatives considered

• More aggressive system prompts ("ALWAYS check memory before acting") — tried extensively, doesn't work under task pressure
• Auto-injecting full MEMORY.md into every turn — too expensive, doesn't scale
• RAG on every message — too noisy, retrieves irrelevant content

The sweet spot is context-aware mandatory retrieval: the system detects what kind of action is about to happen and injects relevant memories before the model proceeds.

Impact

This would be a significant differentiator. Every agent framework has memory storage. None of them solve the "model forgets to check" problem reliably.