openclaw - 💡(How to fix) Fix [Feature]: Parametric consolidation channel for dreaming pipeline (CLS Phase 4) [2 comments, 2 participants]

{
  "dreaming": {
    "phases": {
      "consolidation": {
        "enabled": true,
        "method": "lora",
        "provider": "local",
        "loraRank": 8,
        "minPatternsToTrigger": 5,
        "regressionGuard": true,
        "regressionThreshold": 0.95,
        "checkpointRetention": 5
      }
    }
  }
}


### Alternatives considered

**Alternatives considered:**
Bigger context windows — Does not help. The paper proves (Theorem 1) that the generalization gap is context-window-independent. Larger context holds more notes but cannot synthesize novel concept combinations the model hasn't encoded.

Better retrieval (learned retrieval policies, reranking) — Improves recall quality for seen patterns but does not address unseen compositional combinations. The Ω(k²) coverage requirement persists regardless of retrieval optimization.

Sleep-time text compression (current approach) — OpenClaw's existing dreaming already does this well. But compressing text in an external store produces better-formatted notes; the agent is still a "well-organized novice" with frozen weights. This is explicitly addressed in the paper as insufficient (Section 3.3).

External fine-tuning workflow — Users could manually export dreaming output and fine-tune separately. This is fragile, unsupported, and loses the automated cron-based consolidation cycle that makes dreaming powerful.



### Impact

**Impact:**
Affected: All long-running OpenClaw agents using dreaming/memory-core Severity: High (architectural limitation, not a bug) Frequency: Continuous — every session runs against frozen base weights regardless of accumulated experience Consequence: Agents accumulate months of dreaming output but show zero improvement on novel compositional tasks. Users invest in dreaming configuration expecting the agent to "get smarter" but hit a provable generalization ceiling. The gap between user expectation and system capability grows over time.

### Evidence/examples

**Evidence/examples:**
Paper: "Contextual Agentic Memory is a Memo, Not True Memory" (arXiv:2604.27707) — formalizes the generalization gap with Theorem 1 (Compositional Sample Complexity Separation) and names the Frozen Novice Problem.
Empirical: Yao et al. (2026) ParamMem shows parametric storage outperforms retrieval storage, with the gap growing on novel compositional transfers.
OpenClaw observation: Agent with 3499 recall entries, 18 dream corpus files, and active dreaming (0 5 * * * for weeks) still falls back to base model reasoning on novel multi-concept queries. memory status --deep confirms healthy pipeline; the limitation is architectural, not operational.
Existing infrastructure: OpenClaw already has recall frequency tracking, concept tagging, session corpus ingestion, and cron-based async execution — all prerequisites for consolidation. Only the final LoRA/adapter step is missing.


### Additional information

[consolidation-channel-proposal.md](https://github.com/user-attachments/files/27333326/consolidation-channel-proposal.md)