Persistent memory
Knowledge outlives the session. Insights are written to a file-based vault that's still there tomorrow, next week, and three projects from now.
durable by default
Open source · agent-agnostic memory layer
Your AI agents finally remember.
A durable, file-based knowledge vault that plugs into any coding agent, so the lessons from this session are still there in the next one.
curl -fsSL https://getagentbrain.com/install.sh | bash
The problem
Every new session, your agent starts with amnesia.
It forgets
The fix you found yesterday, the architecture you explained, the convention you agreed on. Gone the moment the context window closes.
It repeats mistakes
The same dead ends, re-debugged from scratch. You become the long-term memory, pasting the same context again and again.
It can't share
Switch from one agent to another and none of the hard-won context follows. Each tool lives in its own silo.
What it does
A memory layer, not another agent.
agentBrain is the durable layer underneath the tools you already use. It captures what's worth keeping, validates it, links it, and serves it back.
Agent-agnostic
One vault, shared across Claude Code, Pi, Copilot CLI and Gemini CLI. Switch harnesses without losing a thing. The knowledge is yours, not the tool's.
one base, every agentSelf-learning
Captures patterns, fixes and decisions as you work. Proven once, written with low confidence; seen again, promoted. Speculation never makes the cut.
capture as you goWiki-linked graph
Notes link to each other with [[wiki-links]]. The result is a connected map of your real architecture, not a folder of orphans.
Skills & hooks
/save-learning, /doctor, /brain-review, peer-review and more. Hooks fire at session start and capture insights without you asking.
Full lifecycle
One-command setup, idempotent updates that never clobber your data, and a clean, symmetric uninstall. It respects your machine on the way out.
install · update · removeAddons & extensions
Opt-in modules: an MCP server, event-bus, web analytics and more. Install what you need; each addon has a symmetric uninstall.
install what you need
One connected mind
Every note, wired into one network.
Learnings, projects, preferences and references become nodes in a single shared graph. This is the shape of a knowledge base that has actually been worked in.
Interactive
Your knowledge becomes a connected map.
Every note is a node. Every [[link]] is an edge. Over time the vault stops being a pile of files and starts being a graph that mirrors how your work actually fits together.
Hover a node to light up its neighbours. Drag one to feel the connections.
How it works
Three steps. Then it gets out of your way.
01
Install
Run the setup script once. It wires agentBrain into every agent it detects and leaves a single Markdown vault on disk.
02
Work normally
Keep coding. Hooks read the vault at session start and capture real learnings as you go, or you save one explicitly with a skill.
03
Agents recall
Next session, on any agent, the relevant notes are already in context. No re-explaining, no re-debugging.
> /save-learning vite needs base:'./' for relative builds
› validating … not a duplicate, actionable, proven
› frontmatter + uuid5 … ok
› linking related notes … 2 found
✓ saved to local/learnings/vite-relative-base.md
> /doctor
› vault healthy · 142 notes · 0 broken links
› validating … not a duplicate, actionable, proven
› frontmatter + uuid5 … ok
› linking related notes … 2 found
✓ saved to local/learnings/vite-relative-base.md
> /doctor
› vault healthy · 142 notes · 0 broken links
---
date: 2026-06-17
type: learning
tags: [vite, build, frontend]
confidence: high
source: session
id: a3f1c9e2-…-uuid5
---
date: 2026-06-17
type: learning
tags: [vite, build, frontend]
confidence: high
source: session
id: a3f1c9e2-…-uuid5
---
Vite relative-path builds
Set base: './' in vite.config when the build is served from a subfolder or opened from the filesystem. Without it, asset URLs resolve from root and 404.
Lifecycle
It respects your machine.
Setup is one command. Updates are idempotent: run them as often as you like, nothing breaks. And the uninstall is symmetric, removing exactly what it added and never touching your knowledge.
Your vault is plain Markdown in a folder you control. No database, no lock-in, no service to trust. Delete agentBrain and your notes are still sitting right there.
Setup
One command detects your agents and configures them. Run setup.sh and you're done.
Daily use
Idempotent updates and a /doctor health check keep the vault honest. Re-run anytime; it converges, never duplicates.
Uninstall
Clean and symmetric. It reverses each config edit it made and leaves your Markdown vault untouched.
Stop being your agent's memory.
Give your tools a brain that persists. Open source, agent-agnostic, and yours to keep.
curl -fsSL https://getagentbrain.com/install.sh | bash