MADB: Local Causal Memory for AI Agents
Give Claude Code, Codex, and MCP-compatible agents durable memory with a traceable record of why they acted. Install in one command. No account. No hosted service. Your agent memory stays on your machine.
Persistent, causal memory that installs in one command.
MADB (Meta-AgentsDB) is a local MCP memory server. It gives Claude Code, Codex, the Anthropic SDK, and any MCP-compatible agent durable memory across sessions, composite-scored recall, and a traceable record of why each decision was made.
Durable memory
Agents stop starting blank. Every action is stored and recalled across sessions, so they build on their own history instead of forgetting it.
persistent memory for AI agentsCausal lineage
Every memory links to what caused it. Walk any decision back to its root with trace_cause — the audit trail behind what an agent did.
Local & yours
Runs on your own machine. Memory never leaves your device; anonymous counts-only telemetry, opt out anytime.
local AI agent memoryOne line. No account. No hosted service.
MADB registers as an MCP server, so it works with any MCP-compatible agent.
claude mcp add madb -- uvx madb-mcp-serveruvx madb-mcp-serverpip install madb-anthropicTwelve stable tools your agent calls automatically.
In Claude Code a bundled skill makes Claude reach for MADB on its own; the same tools work in Codex and any MCP client.
remember
Write a memory event with content, caused_by[], tags[], importance.
recall
Composite-scored semantic recall — query, top_k.
get_memory
Fetch one memory event by id.
search
Broad query across scope.
list_recent
Most recent records, optional tag filter.
forget
Delete a memory — you own and can delete your data.
trace_cause
Walk the causal DAG — answer "why did the agent do X."
save_skill
Persist a reusable skill.
recall_skill
Recall skills by relevance.
list_skills
List persisted skills.
summarize
Rolling summary of memory.
stats
Engine counters.
Auditable AI agents are a precondition for production.
Most agent projects stall before production — not because models can’t act, but because no one can answer for what they did. MADB turns an agent’s decisions into an agent audit trail you can replay on demand.
How is MADB different from a vector DB or chat history?
Vector search finds similar text. Chat history is a transcript. MADB stores why — causal edges between decisions — and recalls on more than similarity alone.