Meridian

The open-source agent OS. Persistent cognitive memory, voice as a first-class channel, batteries-included plugins. By ATERNA AI. Create your legend.

███╗   ███╗███████╗██████╗ ██╗██████╗ ██╗ █████╗ ███╗   ██╗
████╗ ████║██╔════╝██╔══██╗██║██╔══██╗██║██╔══██╗████╗  ██║
██╔████╔██║█████╗  ██████╔╝██║██║  ██║██║███████║██╔██╗ ██║
██║╚██╔╝██║██╔══╝  ██╔══██╗██║██║  ██║██║██╔══██║██║╚██╗██║
██║ ╚═╝ ██║███████╗██║  ██║██║██████╔╝██║██║  ██║██║ ╚████║
╚═╝     ╚═╝╚══════╝╚═╝  ╚═╝╚═╝╚═════╝ ╚═╝╚═╝  ╚═╝╚═╝  ╚═══╝

Create Your Legend.

---

What Meridian is

Meridian is a cognitive agent runtime organized around seven layers: IDENTITY, CONTEXT, SKILLS, MEMORY, CONNECTIONS, VERIFICATION, AUTOMATIONS.

Other agent stacks treat each of those as a discipline you maintain by hand. Meridian materializes them as a typed runtime, with CORTEX cognitive memory wired in at the spine and voice as a first-class channel.

Built by ATERNA AI. Ships under MIT. Powers ATERNA's production agent fleet.

---

What makes Meridian different

	LangChain / Mastra / CrewAI	Mem0 / Letta / Zep	voice-only platforms	Meridian
Category	Orchestration framework	Memory SDK	Voice platform	AgentOS runtime
Cognitive memory	bring-your-own	vector / graph	within-session only	CORTEX-native, in-process
Voice with cross-call memory	no	no	no	yes (voice channel + CORTEX channel:voice valence)
Verification layer	discipline	none	none	runtime-enforced
Per-agent isolation	shared backends	shared backends	shared accounts	dedicated triad: Neon + Voyage + OpenRouter per agent
Bundled plugins	DIY per integration	n/a	n/a	gog / web-search / github / limitless preinstalled
AgentOS portability	partial	none	none	seven-layer filesystem readable by any tool

---

Bundled plugins

Meridian ships plugins as first-class citizens, not "bring your own SDK." Every Meridian agent gets the following the moment you meridian skills install <name>:

Skill	What it does	How
google	Gmail, Calendar, Drive across multiple mailboxes (incl. Workspace DWD)	Bundled gogcli (MIT), auto-downloaded + checksum-verified
web-search	Real-time web search and synthesized answers with citations	Tavily API (free tier, 1000 searches/month)
github	Read repos / issues / PRs, post comments	GitHub REST API + personal access token
limitless	Pull lifelog transcripts from Limitless pendant into CORTEX	Limitless API

Setup for each is one command:

meridian skills install web-search
meridian skills setup web-search   # paste API key, done

Coming next: slack, notion, linear, browser (Playwright bundle). Plugins are MIT-licensed and live under skeleton/SKILLS/<name>/. Copy any one as the template for your own.

---

Install

Requires Node 20+ and pnpm. A CORTEX server reachable on MERIDIAN_CORTEX_URL (defaults to http://127.0.0.1:3100); the open-source CORTEX lives at Rezzyman/cortex and runs alongside Meridian.

git clone https://github.com/Rezzyman/meridian
cd meridian
pnpm install
pnpm link --global   # exposes `meridian` and `mer` on $PATH

The CLI launcher runs straight from src/ via tsx, so no build step is needed for daily use. If you want a compiled dist/ for production deployment, that path is on the roadmap; for now meridian and mer work out of the box.

Bring up CORTEX

Meridian needs a CORTEX server running before any agent can recall or encode memory. The open-source CORTEX is its own repo. Quickest path with Docker:

git clone https://github.com/Rezzyman/cortex
cd cortex
cp .env.example .env
# Edit .env: set VOYAGE_API_KEY (free tier at voyageai.com)
docker compose up -d
npx tsx scripts/run-migrations.ts
npx tsx src/index.ts        # REST API on :3100

Or run Postgres + pgvector yourself instead of Docker; full instructions are in the CORTEX repo. Either way, the result is a CORTEX server listening on http://127.0.0.1:3100 (the default Meridian expects). If you run CORTEX on a different port or host, set MERIDIAN_CORTEX_URL in your agent's .env to match.

---

Quick start

Spin up an agent in five minutes:

meridian init aria               # creates ~/.meridian/aria/
# Edit ~/.meridian/aria/.env and set the three required keys:
#   NEON_DATABASE_URL=...        (dedicated Neon project for this agent)
#   VOYAGE_API_KEY=...           (dedicated Voyage AI key)
#   OPENROUTER_API_KEY=...       (dedicated OpenRouter key)
# Optional channels:
#   VAPI_API_KEY=...             (voice)
#   TELEGRAM_BOT_TOKEN=...       (chat)

meridian doctor                  # validate the foundation

Install plugins:

meridian skills install web-search   # Tavily search
meridian skills install github       # GitHub read + comment
meridian skills install google       # Gmail / Calendar / Drive (auto-downloads gog)
meridian skills setup web-search     # paste your Tavily key
meridian skills setup github         # paste your GitHub PAT
meridian skills setup google         # OAuth multiple Google accounts

Bring channels online and start chatting:

meridian gateway                 # HTTP gateway on :18889 + Telegram + voice
meridian                         # interactive REPL

Add a specialist that inherits the hub's CONTEXT and MEMORY:

meridian init outbound --inherits aria

---

The seven layers

~/.meridian/<agent>/ materializes the AgentOS:

IDENTITY/        AGENT.md, USER.md
CONTEXT/         stakeholders.md, strategy.md, principles.md, ...
SKILLS/          google/, github/, web-search/, limitless/, calendar-prep/, commitment-ledger/, ...
MEMORY/          cortex.config, decision-logs/, relationships/, processes/, episodic/
CONNECTIONS/     mcp.json, calendar.config, inbox.config, slack.config
VERIFICATION/    <skill>.checks.md, audits/
AUTOMATIONS/     dream-cycle.cron, weekly-audit.cron, inbox-scan.cron, ...
config.yaml      .env       state.db       sessions/       logs/

This structure is portable. Any harness that reads markdown can consume a Meridian agent home: Claude Code reads IDENTITY/AGENT.md, OpenClaw reads SKILLS/, Cursor reads CONTEXT/. Meridian is the best runtime for the OS, not the only one.

---

CORTEX-native turn lifecycle

user input
   ↓
preTurn hooks
   ↓
CORTEX recall (CA3 pattern completion against the agent's dedicated Neon DB)
   ↓
recall folded into <cortex_recall> system prompt section
   ↓
provider call via Vercel AI SDK (primary + fallback chain, smart-routing)
   ↓
tool-use loop (built-ins + filesystem skills + MCP)
   ↓
postTurn hooks
   ↓
verification checks (block | warn)
   ↓
CORTEX encode (hippocampal pipeline, valence-tagged, channel-aware)
   ↓
session.append + checkpoint

The dream cycle runs in-process via a Node setInterval worker. No external cron. No "gateway crashed → dream skipped → memory stale" failure mode.

---

Voice with cross-call memory

The headline feature.

Voice assistants in the rest of the market have within-session memory only. Meridian wires the voice channel to CORTEX so every voice transcript is encoded with channel:voice valence, and the next call from the same phone number triggers cross-call recall:

"Hi John, glad you called back. Earlier you were asking about the Oak Hills quote. Did you want to schedule the inspection now?"

That experience is impossible on a raw voice stack without code glue. Meridian gives every voice line a real receptionist's memory.

---

Commands

meridian init <slug>            Seed a new agent home (--template, --inherits)
meridian agents                 List configured agents
meridian use <slug>             Switch active agent
meridian doctor                 End-to-end health check
meridian deploy --intake X.json Run the 20-minute provisioning pipeline
meridian audit                  Run the AgentOS retrospective
meridian gateway                Start HTTP gateway + Telegram + voice channels
meridian                        Open the interactive REPL (default)

Inside the REPL: /help, /cortex, /recall, /encode, /dream, /audit, /skills, /quit.

---

Open-source vs commercial

MIT under this repo:

• Meridian runtime (CLI, REPL, gateway, channels, skills, sessions, verification, audit, automations)
• Seven-layer AgentOS spec
• Voice channel adapter (VAPI) + Telegram channel
• CORTEX V2.4 client bindings (compatible with the public CORTEX server)
• Encrypted vault (AES-256-GCM, per-agent isolated)
• Skills v2 spec + bundled plugin pack (google, web-search, github, limitless)
• DreamWeaver in-process consolidation cycle
• AutomationManager (cron skills with full tool access)

Commercial (ATERNA-licensed):

• Quartz (BSL-1.1): frontier paid memory layer benchmarked at 94.53% on LongMemEval-oracle, statistically tied with the public state of the art. Drops in as a Meridian provider via MERIDIAN_MEMORY_PROVIDER=quartz. Production-deployed inside ATERNA today; available for licensing.
• Operator dashboard (web) and the managed cloud version of Meridian for non-technical operators.

The public release uses CORTEX by default. Quartz lives behind the same MemoryProvider interface, so upgrading is a config flag plus an installed package, not a port. When the Quartz package is missing, the runtime logs a notice and falls back to CORTEX so the agent always boots.

---

Status

v1.2.0, first public release. Reserve a design-partner slot at aterna.ai/meridian.

What's new in 1.2:

• MemoryProvider seam. Pluggable memory backend. CORTEX is the open-source default; ATERNA-licensed Quartz drops in via MERIDIAN_MEMORY_PROVIDER=quartz with graceful fallback when absent.
• Interactive plugin setup. meridian skills setup <name> walks the operator through paste-and-validate flows for Tavily, GitHub PAT, Limitless, and Google OAuth. Masked input. Bad keys never reach the vault.
• Quartz live in production at ATERNA. Arlo (our Chief of Staff agent) now recall-routes through the Quartz pipeline on every turn. Statistically tied with the public LongMemEval-oracle SoTA at 94.53%. Drop-in for any Meridian agent that licenses it.
• Bundled plugins ship 24 tools across google, github, web-search, limitless. Every install is a bundled-binary or paste-the-key flow; nothing requires a separate harness.

What's working today:

• Multi-channel agents (CLI, Telegram, voice via VAPI) with cross-channel memory
• Skills v2 with bundled plugins and OAuth-via-CLI patterns
• Encrypted vault, passphrase-gated tools, voice unlock guard
• Scheduled automations that call tools (e.g. inbox-scan every 30 min)
• Per-agent isolation (Neon + Voyage + OpenRouter per agent)
• Live production fleet at ATERNA: Arlo (Chief of Staff) on dedicated VPS

What's coming next: Slack and Notion plugins, Playwright browser bundle, observation-pipeline integration for Quartz, observability dashboard, public registry.

---

Built on the shoulders of

• @aterna/cortex: the cognitive memory architecture
• Vercel AI SDK: provider abstraction and streaming
• grammy: Telegram bot framework
• fastify: HTTP gateway
• better-sqlite3: session store
• @modelcontextprotocol/sdk: MCP interop

---

License

MIT. © 2026 ATERNA AI.