Solution · Agent memory

Give your coding agent memory that lasts.

Cross-session memory for Claude Code, backed by Vectros. Install a package, wire a few hooks, done. No app code.

Request early accessRead the docs

The alternative

What you'd otherwise build.

Most coding-agent memory today is a single-user local toy, by construction.

  • A local memory store. A folder of files, or a local vector DB, living only on your machine.
  • Manual recall. You decide when to go dig up prior context, instead of it happening automatically.
  • No real isolation. Fine for one project on one laptop; nothing to fall back on beyond that.
  • No audit trail. What got remembered, when, and from which session: not tracked anywhere.

Install

Two commands.

terminal
npm install -g @vectros-ai/claude-code-agent-memory
claude-code-agent-memory init

Recall runs automatically at the start of a turn. Capture distills durable lessons out of a session in the background. Nothing reaches your live knowledge base without your review.

What's actually different

Not a single-user local memory toy.

Built on the same typed records and hybrid search building blocks on this site, so what you get is multi-tenant-capable, isolated, scoped-key-secured, and auditable by default: the same infrastructure a regulated customer app would run on, powering your coding agent's memory instead. See how the access-control mechanics work underneath that. It pairs naturally with the Agentic-SDLC Knowledge Base blueprint: its schema shape and candidate-record queue are what this package is actually designed around, though each works fine on its own too.

What this isn't. Pre-1.0, beta. Scoped to Claude Code today, not a generic any-agent memory tool. Captured candidates aren't filtered for sensitive content before they land in your own private review queue: unreviewed, but never public or searchable until you act on one. Every hook fails open: a missing credential or unreachable API degrades to a no-op rather than breaking your session.

See it in use

We run this on our own engineering org.

This session, and every Claude Code session on this codebase, recalls and captures through this exact package, paired with the Agentic-SDLC Knowledge Base blueprint it's designed around. Not a demo built to show it off: our own day-to-day engineering memory runs on it.

One example of a turnkey product built on the substrate, not a special capability unique to coding agents. See the database and search building blocks it's built from.

View on GitHubView on npm