Self-Driving AgentsGitHub →

Claude Code

Claude Code that learns from every project.

--harness claude-code

What you get

Claude Code in a project directory has no memory of prior sessions there — every conversation starts fresh. Self-driving agents make Claude get sharper over time on each project: it picks up the conventions, the patterns that work, and your corrections, and walks into the next session already knowing them.

  • Learns from every conversation — preferences, decisions, and recurring patterns from each project session feed forward to the next.
  • Builds its own playbooks — the agent maintains long-form pages (a personal wiki) that self-rewrite from what it has learned in this project.
  • One agent per project directory — same agent name in two different project dirs keeps its learning separate. Same dir means same memory.

Setup

  1. 1

    Have Claude Code installed

    The CLI shells out to claude plugin install and writes to ~/.claude/settings.json. Install Claude Code first if you haven't already.

    npm install -g @anthropic-ai/claude-code
  2. 2

    Install the agent

    One CLI call adds the marketplace, installs (or updates) the plugin, configures the connection (interactive on first run), stages seed content, and allowlists the right tools and skill.

    npx @vectorize-io/self-driving-agents install marketing/seo --harness claude-code
  3. 3

    cd into the project directory

    The agent's memory is scoped to the project directory — same agent in a different cd means a different bank. Always start claude from the same project dir.

  4. 4

    Start Claude Code and run the printed prompt

    The CLI printed a one-liner. Pasting it scaffolds a Claude Code subagent, ingests every staged file, and creates initial knowledge pages.

    /hindsight-memory:create-agent marketing-seo from ~/.self-driving-agents/claude-code/marketing-seo

How it works

The integration is the hindsight-memory plugin, installed from the vectorize-io/hindsight marketplace into Claude Code. It registers four hooks: SessionStart runs a health check; UserPromptSubmit does pre-turn recall and injects an invisible <hindsight_memories> block as additionalContext; Stop does post-turn retain (chunked, default every 10 turns with 2-turn overlap); SessionEnd cleans up. Connection lives in ~/.hindsight/claude-code.json (external Hindsight, auto-managed local hindsight-embed daemon, or existing daemon). Tool and skill allowlist is in ~/.claude/settings.json under permissions.allow. The plugin also exposes the tools below.

Tools the agent gets

list_pagesList page ids and names.
get_pageRead the content of a page.
create_pageDeclare a new page recipe.
update_pageChange a page's name or source query.
delete_pageRemove a page.
recallSearch retained memories.
ingestAdd new text content as a memory.
ingest_fileRead a local file and ingest it.
get_current_bankGet the active bank id.

Bank mapping

Resources

Claude Code →hindsight-memory plugin →Hindsight →