Local-first memory for AI agents
Give your agents continuity. Keep the receipts.
Nuzo gives Codex, Claude Code, and MCP-compatible hosts durable context—with explicit writes, bounded recall, and a local audit trail you control.
Select a stage to inspect the boundary.
01 / THE BOUNDARY
Memory should extend an agent.
Not escape your control.
Most memory systems optimize for remembering more. Nuzo starts with a harder question: can you see what was retained, verify its source, change your mind, and remove it completely?
Recall without mutation
SessionStart and UserPromptSubmit are read-only recall hooks. They do not write memory, and recalled records remain untrusted data.
Inference becomes a draft
An agent can propose a useful memory, but the proposal reports memory_writes: false until you confirm the exact content.
Confirmed means accountable
Scopes, provenance, timestamps, revisions, relations, and lifecycle events make each stored record inspectable instead of mysterious.
02 / ONE LOCAL CORE
One memory layer.
Every supported host.
The CLI, MCP server, Codex plugin, and Claude Code plugin share the same core policies and local store. There is no second, hidden memory path.
Explore the architecture03 / START LOCAL
Two commands.
No account required.
nuzo setup detects supported hosts, previews changes, and asks before writing configuration.
npm install --global @nuzo/memory@1.1.0
nuzo setup
Host-specific setup and updates
nuzo setup --codex --yes
nuzo setup --claude-code --yes
nuzo setup --all --yes
nuzo update
nuzo memory manage
04 / PROOF, NOT PROMISES
The contracts are public.
Your memory. Your machine. Your decision.