Openclaw Memory Max

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.

What's Active

Automatic (no action needed)

• Auto-Recall: Before every turn, your most relevant memories are automatically injected into your context as <relevant-memories> XML blocks. You don't need to search — relevant context appears automatically.
• Auto-Capture: After conversations, high-value user messages (rules, corrections, preferences) are automatically captured for the nightly consolidation cycle.
• Compaction Rescue: When the context window is compressed, important content is rescued before it's evicted.
• Episodic Memory: Each session is logged as an episode with timestamps, tools used, and key decisions.
• Sleep Cycle: An in-process scheduler runs maintenance every ~24h — prunes the causal graph, decays stale utility scores, truncates old logs, and writes consolidation context for the next session.

Tools Available

precision_memory_search

Cross-encoder reranked search with utility weighting. Returns the top K most relevant memories.

json

{"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_search

Multi-hop retrieval. Searches once, extracts key concepts from results, searches again with those concepts, then merges everything.

json

{"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_utility

Reinforce a memory that proved useful. Increases its future retrieval priority.

json

{"memoryId": "abc-123", "rewardScalar": 0.2}

Call this after a memory helped you give a correct answer.

penalize_memory_utility

Penalize a memory that caused a hallucination or was irrelevant.

json

{"memoryId": "abc-123", "penaltyScalar": 0.2}

Call this when a retrieved memory led you astray.

memory_graph_add

Log a cause-action-effect chain. Automatically deduplicates against existing chains.

json

{"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_query

Search past experience using semantic matching.

json

{"query": "website not loading", "outcomeFilter": "success"}

Call this BEFORE taking major actions to check what worked or failed in the past.

memory_graph_summary

Get a digest of all learned causal knowledge — success/failure counts, most-frequent patterns, recent outcomes.

json

{}

Useful at the start of a session to bootstrap your awareness.

compress_context

Signal that context compression is needed. Returns what was rescued from the last compaction.

json

{"compression_reason": "context window approaching limit after long debugging session"}

Rules

1. Auto-recall is always on — you will see <relevant-memories> blocks in your context. Use them. Don't ignore injected memories.
2. Reward useful memories — when a memory helps you answer correctly, call reward_memory_utility. This trains the retrieval system.
3. Penalize bad memories — when a memory causes a hallucination, call penalize_memory_utility. This prevents future mistakes.
4. Log causal chains — after significant actions (tool use, decisions, fixes), call memory_graph_add. Your future self will thank you.
5. Check experience before acting — before major actions, call memory_graph_query to see if you've encountered this situation before.
6. Use deep search for complex questions — if precision_memory_search doesn't find what you need, try deep_memory_search which follows concept chains across memories.

Configuration

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.

YAML Rule Pinning (opt-in)

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:

markdown

<!--yaml
rules:
  - weight: 1.0
    constraint: "Never delete production data"
  - weight: 0.5
    preference: "Prefer TypeScript over JavaScript"
-->

Rules with weight >= 1.0 appear as CRITICAL CONSTRAINTs in your prompt. Always obey them.

Security note: Only enable this if you control write access to your ~/.openclaw/memory/MEMORY.md file. Any process that can write to that file could influence agent behavior when pinning is enabled.