SOTA Memory Suite — auto-recall, cross-encoder reranking, multi-hop deep search, causal knowledge graph, episodic memory, and nightly sleep-cycle consolidation.
You have the Memory Max SOTA memory system. It upgrades your memory capabilities far beyond the default memory-core plugin.
<relevant-memories> XML blocks. You don't need to search — relevant context appears automatically.precision_memory_searchCross-encoder reranked search with utility weighting. Returns the top K most relevant memories.
{"query": "deployment configuration", "topK": 5}
Use this when you need to find specific information in memory. More precise than the default memory search — uses a cross-encoder model that reads query + candidate together, not just cosine similarity.
deep_memory_searchMulti-hop retrieval. Searches once, extracts key concepts from results, searches again with those concepts, then merges everything.
{"query": "why did the migration fail last time"}
Use this for complex questions where the answer might be spread across multiple related memories.
reward_memory_utilityReinforce a memory that proved useful. Increases its future retrieval priority.
{"memoryId": "abc-123", "rewardScalar": 0.2}
Call this after a memory helped you give a correct answer.
penalize_memory_utilityPenalize a memory that caused a hallucination or was irrelevant.
{"memoryId": "abc-123", "penaltyScalar": 0.2}
Call this when a retrieved memory led you astray.
memory_graph_addLog a cause-action-effect chain. Automatically deduplicates against existing chains.
{"cause": "nginx misconfigured", "action": "added proxy_pass", "effect": "site loaded", "outcome": "success", "tags": ["nginx"]}
Call this AFTER completing any meaningful action to build your experience database.
memory_graph_querySearch past experience using semantic matching.
{"query": "website not loading", "outcomeFilter": "success"}
Call this BEFORE taking major actions to check what worked or failed in the past.
memory_graph_summaryGet a digest of all learned causal knowledge — success/failure counts, most-frequent patterns, recent outcomes.
{}
Useful at the start of a session to bootstrap your awareness.
compress_contextSignal that context compression is needed. Returns what was rescued from the last compaction.
{"compression_reason": "context window approaching limit after long debugging session"}
<relevant-memories> blocks in your context. Use them. Don't ignore injected memories.reward_memory_utility. This trains the retrieval system.penalize_memory_utility. This prevents future mistakes.memory_graph_add. Your future self will thank you.memory_graph_query to see if you've encountered this situation before.precision_memory_search doesn't find what you need, try deep_memory_search which follows concept chains across memories.All features are controlled via configSchema in the plugin manifest. Users configure these in their OpenClaw settings:
| Option | Default | Description |
|---|---|---|
enableRulePinning | false | YAML rule pinning from MEMORY.md into system prompt. Off by default — must be explicitly opted in. |
enableAutoCapture | false | Automatic capture of high-value user messages to sidecar files. Off by default — opt in if you want persistent message logging. |
enableAutoRecall | true | Automatic memory injection before each agent turn. |
Disabled by default. Must be enabled via enableRulePinning: true in plugin config.
When enabled, users can pin critical constraints into the system prompt by adding a YAML block to MEMORY.md:
<!--yaml