AI is on every board's roadmap — but most agent projects stall before production. Not because the models can't act, but because no one can answer for what they do. These are the questions blocking deployment:
The common thread: agents are trusted with real decisions, but they forget, and they can't tell you why they acted. Memory and traceability are the missing foundation — and that's where MADB starts.
MADB is durable, causal memory for autonomous agents. Every decision is stored with what caused it, so your agents build on their own history instead of starting blank — and you can trace any decision back to its cause. The foundation for AI you can trust in production.
Agents stop starting blank. Every action is stored and recalled across sessions, so they build on their own history. Chief Architect
Every memory links to what caused it. Walk any decision back to its root with trace_cause — the "why" behind what an agent did. CISO · Trust & Risk
Runs on your own machine. Memory never leaves your device; counts-only telemetry, opt-out anytime. Security
This is what's real today: durable memory and a traceable record of why an agent acted. Enforced governance, certification, and tamper-evident audit are the paid layer we're building on top — see the roadmap →
Meta-Agents.AI provides MADB (Meta-AgentsDB) — lightweight, durable, causal memory for autonomous agents. It registers as an MCP server, so it works with any MCP-compatible agent — Claude, Codex, and more. No hosted service, no account, no key.
One line. MADB is live — 13 memory tools, no config.
Or zero-install with uvx madb-mcp-server.
For Anthropic-SDK code — one import swap: from madb_anthropic import Anthropic.
From here, every session gains durable, causally-linked memory. Shown with Claude Code, where a bundled skill makes Claude reach for MADB automatically — the same MCP tools work in Codex and any MCP client (wire them in via your client's config).
Every decision and result is stored as a memory, with context.
Each memory links to what caused it. Ask "why did I do X?" and trace_cause walks the chain back to the root.
Recall is composite-scored — similarity, recency, causal proximity, importance, tag overlap — so the right memory comes back, not just the newest.
MADB (Meta-AgentsDB) registers as an MCP server, so it works with any MCP-compatible agent — Claude, Codex, and more. 13 memory tools, durable on your own disk. No hosted service, no account, no key.
In Claude Code, a bundled skill makes Claude reach for MADB automatically. The same tools work in Codex and any MCP client — wire them in via your client's config.
Or zero-install with uvx madb-mcp-server.
For Anthropic-SDK code — one import swap.
{ "mcpServers": { "madb": { "command": "uvx", "args": ["madb-mcp-server"],
"env": { "MADB_DATA_DIR": "~/.madb/data", "MADB_TENANT_ID": "default" } } } }The stable core Claude calls. Each accepts an optional tenant_id.
Add durable memory to existing Anthropic-API code with one import swap — everything else unchanged.
MADB measures how much, never what: an anonymous, salted install id and counts only — never your memory content, queries, or payloads. Opt out any time with MADB_TELEMETRY=off. Full telemetry detail →
Install, run, and watch your agent start building on its own history.
In Claude Code, one line registers MADB as an MCP server:
No account, no key. Memory is created at ~/.madb/data on first run.
In Claude Code, a bundled skill makes Claude reach for MADB on its own — writing what happens with remember and pulling back what matters with recall. In Codex and other MCP clients, the same 13 tools are available to call.
Because every memory links to what caused it, you can walk the chain back to the root:
Recall is composite-scored across five signals — similarity, recency, causal proximity, importance, and tag overlap — so the right memory comes back, not just the newest.
Your agent now has durable, causal memory it can build on — and recall the why behind any past decision.
MADB's free local tier gives your agents durable, causal memory — recall what happened, and recall why. It's generous, it's local, and it stays free.
We're building a paid and hosted tier on top of the memory engine. These are in development, not shipped today — if any of them matter to you, tell us and help shape them.
This is the layer that turns durable memory into accountability — the reason teams running autonomous agents in production will want it. The free tier is where memory lives; this is the layer being built on top.
Tell us which of these matter for your deployment — design partners help set the direction.
MADB runs locally on your machine. Your memory — everything your agent stores and recalls — never leaves your device. To understand adoption, the MADB MCP server includes a small, anonymous usage beacon. Here's exactly what it does, and how to turn it off.
All memories live in your local data directory (~/.madb by default). MADB never uploads, syncs, or transmits your memory content anywhere. There is no account and no server holding your data.
The beacon reports anonymous counts and environment only — never your data.
The install ID is a random UUID stored at ~/.madb/.install_id. It is not derived from your username or hostname and contains no personal information. It exists only so two heartbeats from the same install aren't double-counted.
The beacon is fire-and-forget on a background thread with a short timeout. It never blocks a tool call and never interrupts your work — if the network is down, it simply gives up silently.
Anonymous counts tell us whether agents are actually reaching for MADB and where installs get stuck — so we can make it better. That's the whole purpose. We've designed the beacon so it can never carry your data, by construction.
Questions? Email 01@meta-agents.ai.
trace_cause.~/.madb by default. It never leaves your device — there's no account and no server holding your data.MADB_TELEMETRY=off. Full detail →forget tool removes a memory; the store lives on your machine, so it's yours to manage.Agents that forget the moment a session ends start every task from blank. MADB is the memory they keep — durable, causal, and yours.
Meta-Agents.AI builds MADB (Meta-AgentsDB): lightweight, durable, causal memory for autonomous agents.
It runs locally, free, on your own machine. Every memory links to what caused it — so an agent can recall not just what happened, but why. Recall is composite-scored across five signals, so the right memory comes back, not just the newest. It works with any MCP-compatible agent via the Model Context Protocol.
That's the whole idea: give your agents a memory that grows with them, and the ability to recall the reasoning behind any past decision.
Questions, higher limits, or shaping the roadmap — we read every message.