openclaw-memory-max_skill

Overview

This skill implements the Memory Max system: a state-of-the-art memory suite that improves recall, consolidation, and experience-driven decision making. It combines auto-recall, episodic logging, nightly consolidation, and advanced retrieval tools to keep relevant context available and to evolve memory utility over time. Use it to maintain long-term, actionable memory across sessions and automate memory hygiene.

How this skill works

Relevant memories are automatically injected into your context before each turn as structured <relevant-memories> blocks so you can leverage them without manual searching. Advanced tools provide precision cross-encoder reranking, multi-hop deep search, causal graph logging and querying, and utilities to reward or penalize memory utility. A nightly isolated consolidation agent compacts memories, prunes the causal graph, and decays stale scores into a persistent MEMORY.md.

When to use it

• When you need precise recall of past decisions, rules, or user preferences before responding.
• To find distributed information that requires multi-hop reasoning across several memories.
• Before taking major actions to check prior successful or failed approaches via the causal graph.
• After a helpful retrieved memory to reinforce its future priority.
• When a retrieved memory led you astray, to penalize it and reduce recurrence.

Best practices

• Rely on injected <relevant-memories> — they are curated for each turn and reduce hallucination risk.
• Use precision_memory_search for targeted lookups and deep_memory_search for complex, cross-linked queries.
• Call memory_graph_add after meaningful actions to grow the causal knowledge base.
• Reward memories that produce correct outcomes and penalize those that cause errors to tune retrieval.
• Review memory_graph_summary at session start to bootstrap situational awareness.

Example use cases

• A troubleshooting flow: query memory_graph_query for past similar incidents, then apply the previously successful fix and log the result with memory_graph_add.
• A long-term user assistant: auto-capture preferences and personal details, then rely on auto-recall to personalize future responses.
• Complex incident postmortem: run deep_memory_search to collect scattered notes and produce a consolidated incident summary.
• Context-heavy sessions: use compress_context to preserve critical rescued content when the context window approaches its limit.
• Nightly consolidation: let the sleep-cycle agent distill captured rules and prune noisy or stale memories for a cleaner knowledge base.