CALIBER Dashboard
AI Agent Memory Management — Pack Editor, Forensics, Policy Visualization, and Memory Stream
About CALIBER
Agents are state machines on event DAGs. Deploy them here.
CALIBER is a Postgres-native memory framework for AI agents. Agent memory is an engineering problem, not a retrieval problem. Every decision your agent makes is traceable. Every mutation is auditable. Every state transition is recoverable.
Your agent's memory is a DAG of immutable events, hash-chained with Blake3, validated through a pure-function gate pipeline, and stored in PostgreSQL as the source of truth.
Dashboard Features
The CALIBER Dashboard is a web application for managing and inspecting AI agent memory. It requires JavaScript to run.
- Pack Editor: Visual editor for agent configuration packs. Define agents, providers, memory profiles, toolsets, and policies using the cal.toml DSL with Markdown prompt files. Syntax validation, live preview, and intelligent suggestions.
- Event DAG Forensics: Inspect the immutable event history of your agents. Every state change — creating artifacts, recording decisions, closing scopes, delegating to other agents — is a Blake3 hash-chained event in a directed acyclic graph. Trace causal ordering, explore branching from multi-agent fan-out, and reconstruct state at any point in time.
- Policy Pipeline Visualization: Trace mutations through the 4-stage pipeline: Assemble (I/O, fetching context) → Gates (pure computation, deterministic policy checks) → Commit (atomic transaction) → Receipt (Blake3 cryptographic proof). See exactly which gates passed or failed and why.
- Memory Stream: Real-time view of agent memory across all four projections — episodic (what happened), semantic (what the agent knows), procedural (how to do things), and working (what is happening now). All projections derive from the same Event DAG.
- AI Assistant: Guided agent configuration powered by CopilotKit. Get help writing agent prompts, configuring memory profiles, and setting up multi-agent coordination.
- Dashboard Overview: Agent statistics, trajectory monitoring, scope lifecycle tracking, and system health metrics.
Core Capabilities
- Hierarchical Memory: Tenant → Trajectory → Scope → Turn. Artifacts and Notes survive scope boundaries. Token budget boundaries at every level.
- Mutation Pipeline: All writes flow through Assemble → Gates → Commit → Receipt. Gates are pure computation — no I/O, no side effects, fully auditable.
- PCP (Persistent Context Protocol): Harm reduction layer enforcing dosage limits, contradiction detection, summarization triggers, and recovery checkpoints.
- Multi-Agent Coordination: Typed locks, task delegation, agent handoffs, inter-agent messaging, and BDI model for agent state.
- Deterministic Scoring: 6-factor weighted formula — vector similarity, warmth, recency, abstraction level, graph distance, keyword match.
- Event Forensics: Blake3 hash-chained immutable event DAG with complete audit trail and tamper evidence.
Framework Integrations
CALIBER provides persistent memory for any agent framework:
- LangGraph: CaliberCheckpointer for state persistence.
- CrewAI: Crew-level and agent-level memory backend.
- Mastra: Storage adapter for Mastra's memory system.
- PydanticAI: Context provider for PydanticAI agents.
- Vercel AI SDK: Middleware for persistent conversation state.
- CopilotKit: State management backend.
- AG-UI / MCP: Native protocol adapters built into the SDK.
Portability
Your agent is a folder. cp -r my-agent/ /anywhere/. Zero vendor lock-in.
Get Started
git clone https://github.com/heyoub/caliber.git
cd caliber
docker-compose -f docker/docker-compose.yml up -d
curl http://localhost:3000/health