Give your agent a memory. Make it show receipts.
Your agent stops starting from zero — and its word stops being the evidence. Attune carries decisions, bugs, and lessons across sessions, and verifies every change with probes re-run independently of the agent that claims it finished.
Or install it as a Claude Code plugin — no Python required.
Your AI is smart.
Your project has amnesia.
A coding agent spends hours learning your architecture, your constraints, your conventions, the decisions behind the code, and the approaches that already failed. Then the session ends — and most of that knowledge is gone. The next session, the next tool, the next model: each one starts rediscovering your project from zero.
Attune makes that knowledge persistent.
One reliability loop
Every change moves through the same five stages — from a requirement you can point at to an implementation you can trust.
- 01
Specify
Socratic spec engine: requirements, design, and tasks with an approval gate.
- 02
Ground
RAG retrieval cites your code so the agent doesn't invent APIs.
- 03
Build
The agent executes the spec'd tasks — 21 workflows: review, tests, bug prediction, refactor.
- 04
Remember
Cross-session memory and a lessons corpus surface what worked before.
- 05
Verify
Fact-check generated content: imports, flags, links, counts — all real.
Your agent stops starting from zero.
Findings from each session are stashed and recalled in the next — automatically, or on demand with /recall. What persists:
- Architectural decisions
why the code is shaped the way it is
- Coding conventions
the patterns your reviewers actually enforce
- Discovered constraints
the limits that are not written down anywhere
- Lessons from failed approaches
so the same dead end is not explored twice
- Implementation details
the seams, gotchas, and load-bearing quirks
- Project-specific knowledge
domain terms, priorities, and context
Local-first by default — memory lives on your machine, no cloud required. An optional Redis tier adds semantic recall with local Ollama embeddings.
One project brain. Multiple agents.
Attune is the persistent layer beneath your coding agents, not a feature of one of them. Claude Code, Codex, and Antigravity work against the same project memory, specs, lessons, and repository state — handing off work with git-verified packets and giving each other second opinions on real diffs.
Measured, not promised.
pip install attune-ai is the whole stack
The engines below ship inside the one package — no assembly, no extra keys. Each one earns its place with a benefit you can point at.
attune-ragbundledAnswers cite your code
The retrieval engine grounding every generated claim in your actual source — mean faithfulness 0.97, CI-gated.
attune-verifybundledHallucinations caught before merge
The fact-checker: imports import, CLI flags exist, links resolve, counts match — checked against source-of-truth.
attune-redisbundledMemory served fast, owned by you
Serves your git-tracked memory corpus from local Redis; the files stay the store, and everything degrades gracefully without a server.
attune-formsbundledQuestions become one tap
The decision, pushback, and progress cards — structured questions validated on the way back, on every surface.
attune-helpoptionalProgressive help for your own project
Optional standalone reader: concept → task → reference depth for docs generated from your codebase. pip install attune-help.
Inspect everything.
Attune is open source under the Apache License 2.0. Every workflow, every memory operation, every verification step is code you can read, modify, evaluate, and use — commercially included. No black box between your agent and your repository.
View the source on GitHubGet started
Recommended: install both. Free to start, no API key needed — a key only gates direct CLI workflow runs.
In Claude Code
Works standaloneNo Python environment required. Runs on your existing Claude subscription — then type /spec or /coach to start.
Python package
Adds CLI + MCPUnlocks the CLI, the MCP server, and multi-agent workflows. With ANTHROPIC_API_KEY set, large modules fall back to the Anthropic API — API usage bills API credits, separate from a Claude subscription.
Project memory, retrieval, and the help system are local-first — no API key, no cloud required.
Your coding agent shouldn't forget yesterday.
Give it a persistent project memory.