Arshis Memory Pro

Problem

OpenClaw long-term memory often fails by injecting noisy context: low-value memories crowd the prompt, repeated fragments interfere with reasoning, and simple append-only logs lack offline distillation or local retrieval. This plugin treats memory as a local, retrievable, decayed, and evolvable knowledge base rather than a raw conversation log.

How It Works

• Hybrid retrieval: uses LanceDB for 1024-dimensional semantic vectors, combines BM25 keyword search with reranking, and grades memories from 0.1 to 1.0 by importance.
• Dreaming engine: Light Sleep summarizes recent conversations, Deep Sleep extracts long-term patterns, and REM generates cross-category connections; low-value memories decay after 30 days, while high-value, frequently recalled memories are promoted.
• Injection control: defaults to summary-only, injects at most 8 memories per query, uses a 0.72 relevance threshold, and caps memory context at 12,000 tokens.
• Self-evolution: collects retrieval feedback, adjusts weights, decay cycles, and thresholds, and generates strategies from error patterns.
• Modes: primary takes over the memory slot and replaces memory-core; helper coexists with native memory as an auxiliary retrieval and scoring layer.

Boundaries

Best for OpenClaw setups that need local privacy, inspectable storage, and hard context limits. It makes no external network requests, does not call external LLMs, and requires no API keys, but depends on LanceDB and rank-bm25. Auto-capture is disabled by default, and destructive operations require confirmation; enabling primary mode replaces the original memory plugin.