Hybrid recall
BM25 lexical matching fuses with semantic vector similarity via Reciprocal Rank Fusion across 11 recall paths — exact identifiers and fuzzy meaning in one query.
OpenClaw support
OpenClaw memory plugin, backed by libravdbd.
@xdarkicex/openclaw-memory-libravdb replaces the lightweight default memory path with scoped session, user, and global memory owned by the local daemon. Registers as both a memory and context-engine plugin in the OpenClaw gateway.
Cognitive signal filters
Memories tagged as identity, fact, preference, constraint, decision, or episode. Search by kind for surgical recall instead of undifferentiated nearest-neighbor.
Predictive context
The daemon pre-computes what the agent is likely to ask next and injects relevant memory before the turn executes — no extra round-trips.
Contradiction detection
Regex-based anchor extraction with Jaccard dedup automatically marks superseded facts. New information replaces stale claims without manual cleanup.
Half-life decay per kind
Constraints and identity never decay. Facts decay over 180 days. Preferences over 365. Each memory kind follows its own retention curve.
Deterministic state skeleton
L8 extraction pulls decisions, constraints, and next steps via pure heuristics — no LLM call, no token cost, no latency spike.
Markdown ingestion
Watches directories including Obsidian vaults and syncs notes into daemon memory. Agents stay aligned with your written knowledge base.
Multi-tenant isolation
Per-agent isolated vector databases inside a single daemon process. Assign each agent a distinct tenant ID for clean separation.
openclaw plugins install @xdarkicex/openclaw-memory-libravdbOpenClaw >= 2026.3.22 Node.js >= 22 libravdbd daemon Unix socket or TCP/TLS HMAC nonce auth Compaction provider
GitHub — openclaw-memory-libravdb · Talk about OpenClaw integration