Vectors are not enough
Embedding-only recall drops exact identifiers, file paths, flags, stale decisions, hard constraints, and the causal path that made a result useful.
Local-first agentic memory - daemon available now
A cognitive memory kernel for agents that ship.
LibraVDB Memory gives serious agent teams a local daemon for scoped recall, causal continuity, constraint protection, predictive context, and operator-owned memory lifecycle. It is not a prompt trick, not a vector wrapper, and not a cloud dependency.
No telemetry. No phone home. No required LLM, reranker, hosted embedder, or API key.
Daemon available now
Run the local cognitive memory daemon today.
libravdbd ships as a compiled static binary through apt, AUR, Chocolatey, and Homebrew. It runs as operator-managed infrastructure; source code is closed source and protected by the LibraVDB Memory LICENSE and EULA.
apt AUR Chocolatey Homebrew
Host support
Use the daemon from the agent stack you already run.
LibraVDB Memory keeps recall, compaction, graph maintenance, and lifecycle state inside libravdbd. Host integrations stay thin and explicit.
MCP bridge
Tool-native memory access for search, remember, recall context, constraints, decisions, export, forget, status, and journal workflows.
MCP supportOpenClaw plugin
A TypeScript host adapter that replaces the default memory path with daemon-backed assembly, compaction, lifecycle hooks, and scoped recall.
OpenClaw supportHermes provider
A native Python memory provider and context engine with direct daemon gRPC, async turn sync, predictive context, and cross-session recall.
Hermes supportLive memory test
See LibraVDB Memory running in public.
Join the Discord and visit #bots-everywhere, where community-run bots stay connected around the clock with LibraVDB Memory behind them. Tell a bot a fact, come back sessions later, and test whether durable recall survives real conversation.
24/7 bots global testers cross-session recall
Join the DiscordContinuity is the difference between a demo and a system.
LibraVDB Memory keeps facts, constraints, decisions, and temporal context connected across turns and sessions, so agents can carry what matters forward without treating memory like a pile of nearby chunks.
Context needs scope
Session turns, durable user facts, global rules, summaries, and current working context need distinct boundaries instead of one blended transcript pile.
Realtime agents need local paths
Robotics and autonomous runtimes cannot put every recall decision behind a remote memory service and still protect response time.
Useful recall needs credit
A serious memory kernel must learn which prior facts enabled the useful result, not only which item happened to be returned.
The cognitive kernel
Not a plugin. A cognitive memory kernel.
libravdbd owns the hard parts of memory: what to recall, what to protect, what to compact, what to forget, and what to reinforce. Host adapters stay thin; the daemon carries the cognitive load.
Recall
Exact and semantic memory in one path
Identifiers, file paths, constraints, decisions, and natural-language context are retrieved together instead of forcing agents to choose between brittle keyword search and fuzzy vector recall.
Control
Homeostatic retrieval control
TSK fuzzy controllers adapt retrieval, compaction pressure, contradiction handling, authority balance, exploration, and habit suppression from live memory signals.
Causality
Memory that knows why it mattered
Useful recall reinforces the constraints, decisions, and causal paths that made the result possible, so the system learns more than access frequency.
Compaction
Compression without amnesia
Long sessions are compacted with protected anchors, rules, decisions, constraints, and recent working context preserved, instead of being flattened into lossy summaries.
Reasoning
Symbolic protection for serious instructions
Rules, prohibitions, permissions, identity facts, and hard constraints are treated as protected operational state, not decorative text inside a summary blob.
Consolidation
Important memory stops behaving like cache
Repeated usefulness strengthens durable memory and its supporting causal context, while low-value noise can decay, compact, or be removed under operator control.
Deployment
Local-first by default, hosted when you want it
Run the daemon where the agent runs for response-time-sensitive workloads, or join the managed-service waitlist when hosted operations are the right tradeoff.
Operations
Memory belongs in infrastructure
Status, health, export, forget, journal, tenant isolation, and lifecycle hooks give teams a memory runtime they can operate instead of a hidden SDK side effect.
Hybrid recall core
Semantic recall, BM25-style exactness, graph proximity, temporal comparison, and optional second-pass precision work together for technical memory.
Causal graph memory
Memory is scoped state connected by why, how, temporal, and associative relationships instead of flat transcript fragments.
TSK homeostatic control
Fuzzy controllers tune scoring, compaction pressure, contradiction gates, authority, hop decay, information gain, and habit penalties from live signals.
Multi-timescale scheduler
Micro credit, meso structure learning, and macro consolidation let memory stabilize while the agent keeps working.
Counterfactual credit
Useful recall reinforces the decisions, constraints, and ancestor paths that made the result possible.
Synaptic consolidation
Stability signals, engagement credit, and proactive sweeps keep important memory from behaving like disposable cache.
Runtime architecture
The agent host stays thin. MCP bridges, OpenClaw plugins, or Hermes providers pass lifecycle events into libravdbd. The daemon owns recall, inference, graph maintenance, compaction, persistence, and lifecycle control behind a narrow operational boundary.
Agent / Robot HostAdapter / MCPlibravdbdHybrid Vector Graph StoreCognitive Scheduler
Private release
Be first to the hosted service.
The local daemon ships now through apt, AUR, Chocolatey, and Homebrew. The hosted managed version is in private release; join the waitlist for early access.