Local-first memory for OpenClaw that stays inspectable, trustworthy, and useful over time.
Hybrid Memory gives your OpenClaw agent persistent memory without turning memory into a black box. It can remember across sessions, show its work when you need proof, and stay under your control.
Get Started Common tasks Scenarios & benefits View on GitHub
What to remember Why it matters Local-first by default Memory stays on your machine unless you deliberately choose otherwise. Inspectable and controllable You can verify, search, export, back up, and delete it. Hybrid retrieval Structured facts and semantic recall work together, so results are stronger than session-only or vector-only memory.
Without hybrid memory With hybrid memory Every new thread starts from zero Preferences, decisions, and facts carry over You paste the same context again Relevant memories can load automatically each turn Hard to find “that thing we said” Search by meaning and by structure Memory becomes a junk drawer Decay, tiering, and jobs keep storage manageable Hard to tell why a memory showed up Search, verification, provenance, and docs make recall inspectable Lost work if a process crashes badly Durable writes designed for safe recovery
How the engines work (SQLite, vectors, merge, WAL): How it works and Architecture . Narrative examples: Scenarios & benefits .
Document Description Architecture Four-part hybrid architecture, workspace layout, bootstrap files Architecture Center Architecture anchor: core runtime boundary vs adjacent subsystems How It Works End-to-end flow of memory capture, storage, and recall Retrieval Modes Ownership and contracts for interactive vs explicit/deep retrieval paths Memory Protocol Paste-ready AGENTS.md block
Document Description Features Overview Categories, decay, tags, auto-classify, source dates Credential Vault Opt-in encrypted storage for API keys, tokens, passwords Session Distillation Extracting durable facts from old conversation logs Graph Memory Graph-based fact linking and spreading activation WAL Crash Resilience Write-ahead log design and recovery Reflection Layer Pattern synthesis from accumulated facts Conflicting Memories Classify-before-write, supersession, bi-temporal Memory Tiering Hot/warm/cold tiers, compaction, session-end hooks Memory Scoping Global, user, agent, session scope; multi-agent; scope promote Search improvements (RRF, ingest, query expansion) RRF fusion, ingest-files, query expansion Multilingual language keywords Auto-capture and categories in multiple languages; auto-build and build-languages Automatic Categories Category discovery from “other” facts Dynamic Derived Data Index: tags, categories, decay, and other derived data Procedural memory Procedure tagging, recall, auto-skills, promotion gates Skill pipelines Crystallization vs procedure promotion, lifecycle states, operator playbooks Cost optimization playbook Cheap-first tiers, distill guardrails, weekly auditing
┌─────────────────────────────────────────────────┐
│ OpenClaw Agent │
├─────────────────────────────────────────────────┤
│ Hybrid Memory Manager │
│ ┌──────────┐ ┌──────────┐ ┌──────────────┐ │
│ │ SQLite + │ │ LanceDB │ │ Hierarchical │ │
│ │ FTS5 │ │ Vectors │ │ Files │ │
│ └──────────┘ └──────────┘ └──────────────┘ │
│ ┌──────────┐ ┌──────────┐ ┌──────────────┐ │
│ │ WAL │ │ Graph │ │ Reflection │ │
│ │ Layer │ │ Memory │ │ Layer │ │
│ └──────────┘ └──────────┘ └──────────────┘ │
└─────────────────────────────────────────────────┘
# 1. Install the plugin
openclaw plugins install openclaw-hybrid-memory
# 2. Apply recommended config
openclaw hybrid-mem install
# 3. Restart and verify
openclaw gateway stop && openclaw gateway start
openclaw hybrid-mem verify
openclaw hybrid-mem status
openclaw hybrid-mem dashboard