Meridian is a multi-agent platform built on persistent memory, governed behavioral change, and coordinated agent dispatch.
Our Premise
Foundation models are widely available. The differentiation comes from the layers built around them: memory that persists across sessions, governance that keeps behavior coherent over time, coordination that handles the parallelism that complex tasks require.
Meridian is our answer to what that infrastructure looks like in practice. These are the three principles it's built on.
A context window is retrieval. Memory is accumulation. AI systems that accumulate structured experience over time behave differently from systems that reset with every session. The architecture that enables this is where most of the implementation work lives.
Systems that run continuously evolve whether you govern that evolution or not. Governance works as an infrastructure layer, proposal, review, version, audit, rather than a correction applied after drift becomes visible. Systems built with this structure tend to stay coherent at scale.
Capability in complex tasks comes from coordinated agents, not from a single model working sequentially. The task graph, how work decomposes, routes, executes in parallel, and reconciles, is the structural unit of a multi-agent system. Coordination infrastructure is where most of the open engineering problems are.
Platform
Four capability layers. Each independently useful. Together, the infrastructure for autonomous AI systems that actually work at scale.
Parallel task execution with write-back guarantees.
Synthetic agent interaction for testing and growth.
Controlled, auditable evolution for deployed agents.
Contextual memory and inter-system coordination.
Model-agnostic. Runs on Ollama, OpenAI, and Anthropic. Stack: Python · SQLite · React · FastAPI · MCP.
System State
Meridian runs continuously across a multi-node environment. These are measurements from the live system.
A task submitted through the dispatch layer reaches completion, including agent execution, write-back, and status transition, in under 22 seconds. Verified against a live taskboard with full activity log.
TSK-115 · todo → in_progress → review · activity log confirmed
Agents in Meridian carry versioned configuration, behavioral memory that accumulates across sessions, and an explicit growth framework: proposals, review, approval. Task agents are spawned on demand and closed on completion. Governed agents persist and evolve.
persona-core · agent-core · versioned configs · behavioral governance
Five MCP integrations expose real-time ecosystem state to agents: taskboard, doc-source, dev-launcher, agent-core, and dep-scanner. Agents operate with current structured context drawn from 20 live services, not static prompts.
mcp-hub · dev-launcher · taskboard · doc-source · agent-core · dep-scanner
Agent governance cycles, session review, goal discovery, self-model synthesis, and growth proposal generation, run in production. Proposals accumulate in a review queue; approved changes version the agent configuration.
Growth governor · executor · version bump · proposal queue · audit trail
Active research lab · Windows + Mac + VPS multi-node deployment
Built on Meridian
Systems grown inside the Meridian ecosystem, each one a real deployment that stress-tested a different layer of the platform.
Autonomous trading bot built on a six-agent event bus, each agent owning a distinct layer of the pipeline from signal generation to risk arbitration to live execution.
A home network analysis and firewall rule laboratory, live packet capture, device discovery, and guided security exercises, with every privileged operation explicitly permissioned and fully audit-logged.
A Jira-inspired kanban board built for solo developers and AI agents, dual-auth design lets browser sessions and API key–authenticated agents read and write tasks from the same board.
Research & Writing
Thinking in public about autonomous systems, infrastructure, and the parts of production AI that the field is still working through.
AI safety gets discussed at the model level. Agent governance is a different problem, one that almost nobody is building for, and that every team deploying AI at scale will eventually need.
Running a single powerful model on a complex task is an architectural mistake. The systems that will matter are built to exploit parallel decomposition, not work around its absence.
Most organizations deploying AI are optimizing the wrong layer. The frontier isn't the model, it's what surrounds it.
Get in Touch
We're selectively open to research partnerships, collaboration inquiries, and conversations with investors who think about AI infrastructure seriously.
No pitch deck required.
Tell us who you are and what you're thinking about. We'll have a direct conversation about what Meridian is, where it's going, and whether there's a fit worth exploring.