openclaw - 💡(How to fix) Fix Proposal: explicit correction staging for OpenClaw memory after Dreaming [1 comments, 1 participants]

Before Dreaming, I was using memory-governor to explore a broader memory-hardening problem:

• what should count as memory
• where it should go
• what should remain short-term
• what should be allowed to harden into durable rules

The 4.5 Dreaming update changes that picture in a good way.

OpenClaw Dreaming already improves background consolidation from short-term memory into long-term memory. That means a meaningful part of the original problem is now better handled by core, especially daily_memory -> long_term_memory.

What still feels missing is a lightweight layer for explicit corrections and first-sighting emerging lessons:

• things the user corrected once
• things that look reusable, but are not yet proven
• things that should not harden directly into durable rules

Because of that, I do not think the whole memory-governor skill should be absorbed into core as-is. Instead, I think the Dreaming update makes the upstream ask smaller and clearer:

• keep Dreaming as the consolidation engine
• consider whether a few explicit correction-staging ideas are still worth absorbing upstream

I built memory-governor as a community skill to explore that remaining gap.

So this proposal is not "please replace Dreaming." It is:

• Dreaming solved a meaningful part of the old problem
• the remaining gap is now narrower
• explicit correction staging may still be a useful missing piece

The Gap

There is still an awkward space between:

• same-day memory / short-term traces
• reusable durable lessons
• system-level rules

Without an explicit staging layer, explicit corrections tend to become either:

• over-hardened too early
• or buried inside daily notes

Dreaming is already strong at background consolidation and promotion from short-term memory. That is a different job from explicit correction hardening.

Proposal

1. Add an explicit correction staging concept

Something like learning_candidates:

• explicit corrections go here first
• first-sighting emerging lessons go here first
• promotion into reusable lessons stays manual or review-gated

This helps distinguish:

• "the user corrected this once"
• from
• "this should now be treated as durable reusable guidance"

2. Split promotion authority clearly

Recommended split:

• Dreaming-preferred: daily_memory -> long_term_memory
• Manual-only: learning_candidates -> reusable_lessons reusable_lessons -> system/tool/system-style rules

This avoids duplicate promotion authority and keeps Dreaming focused on consolidation.

3. Clarify Dreaming artifact boundaries

It would help to make the following explicit in docs or integration guidance:

• DREAMS.md is a Dreaming artifact
• memory/.dreams/ is Dreaming engine state
• neither should be treated as a standard memory target class

That would reduce confusion for advanced users building host-specific memory workflows, plugins, and skills.

Why This Complements Dreaming

Dreaming is already good at:

• background consolidation
• multi-stage ranking
• explainable promotion from short-term memory

This proposal targets a different gap:

• explicit correction staging
• manual hardening boundaries
• preventing one-off corrections from becoming durable truth too early

In short:

• Dreaming handles background memory consolidation
• explicit correction staging handles manual hardening

What This Proposal Does Not Ask For

This is not asking OpenClaw to:

• absorb the entire memory-governor skill into core
• replace Dreaming with a manual review system
• model DREAMS.md as a user-facing target class
• let background consolidation write directly into system-governance files

Community Reference

I have a working community implementation in memory-governor that currently tests:

• explicit correction staging
• candidate review workflow
• Dreaming-vs-manual promotion authority split
• explicit DREAMS.md / memory/.dreams/ boundary guidance

Happy to share a minimal contract-level diff if that is useful.

Smallest Potential Upstream Pieces

If only the smallest parts are worth considering, the best candidates seem to be:

1. explicit correction staging
2. clearer Dreaming vs manual promotion authority
3. clearer Dreaming artifact boundary guidance