Jossue AlcaláCreate · Build

Own product · Memory for agents

Shared memory for your coding agents,stored in your repository.

AHP+ keeps the project’s state, decisions and test results in a .ahp/ folder, with Git. If you switch from Codex to Claude Code, Cursor or OpenCode, or change account or machine, the next agent reads the same thing and checks project, branch and commit before touching anything.

Talk about my agent workflowSee every command
npx @jossuealcala/ahp-plus@1.4.1 setup .
Demo · 36 s
0
third-party dependencies: just Node and Git
64
command forms, in 16 categories
1.4.1
current version, Apache-2.0 license

Free · Open source · 1.4.1

What it solves

When you change agent, account or machine the project context is lost

A coding agent can follow a conversation, but the conversation does not say which commit is the right one, which decisions still stand or which tests ran. AHP+ writes it down in the repository, where your project already lives.

  • You switch agents
    Without AHP+
    You explain the project from scratch to the next one.
    With AHP+
    It reads .ahp/ and picks up with the current state, the decisions in force and what was already tested.
  • Knowing what was decided
    Without AHP+
    The decision stayed in a chat you no longer open.
    With AHP+
    Every decision, task, bug, risk and piece of evidence is its own file in the repository, with its history in Git.
  • The commit or branch changed
    Without AHP+
    The new agent works on another version and nobody notices.
    With AHP+
    Before continuing it compares project, branch, commit and local changes. If they do not match, it stops and asks you to sort out the difference.
  • You change machine or account
    Without AHP+
    The context stays on the previous machine or account.
    With AHP+
    It travels with the repository. You pull and the next agent has the same thing.

How it works

How work passes from one agent to another

Pick a step to see what is written to .ahp/ and what the receiving agent checks. The commands are the real ones; the identifiers and the commit in the example are illustrative.

What is included

What it stores in .ahp/ and what it installs in your project

  • State, decisions and testsOne file per record (decision, task, bug, risk, QA, requirement, evidence) inside .ahp/, versioned with Git.
  • Hand-offs that get checkedCompares project, tree, branch, commit and local changes. If they do not match, the result is RECONCILIATION_REQUIRED.
  • CheckpointsMark where a session stopped and what the next action is, so you can resume it. They do not replace your commits or backups.
  • Messages between agentsmessage send, reply and inbox leave notices with a SHA-256 fingerprint, which proves they were not altered. They do not store the full conversation.
  • No sensitive actionsIt never commits, pushes, merges, deploys, publishes or deletes. It does not give itself permissions either.
  • Per-agent adaptersCodex with the $ahp skill, Claude Code with CLAUDE.md, Cursor and OpenCode with /ahp, and any other agent with AGENTS.md.

Control and limits

What AHP+ does with your repository and what it does not do

What it does

  • Writes to .ahp/ and stops writing if the commit or the state is no longer the expected one.
  • Installs adapters in plan mode first, without silently overwriting anything.
  • Labels each claim with its certainty: verified, confirmed by you, inferred, unverified, stale or conflicted.
  • Runs on macOS, Windows and Linux with Node 20 or later; it needs no third-party dependencies.

What it does not do

  • Commit, push, merge, deploy, publish or delete.
  • Copy the full conversation or the model’s reasoning.
  • Replace your tests or a person’s review: it only records what was observed.
  • Grant permissions: READY means the hand-off matches the repository, not that anything is authorized.
  • Store tokens or .env files: .ahp/ travels with Git, treat it like the repository.

Who it is for

When it makes sense and when it does not yet

It makes sense if

  • You work with coding agents from an IDE and often switch agent, account or machine.
  • You want decisions and test results to stay in the repository and not in a chat.
  • You have Node 20 or later and a Git repository with at least one commit.

Not yet, if

  • You look for vector memory, an agent orchestrator or a permissions system: AHP+ is none of those.
  • You need it to undo changes for you: checkpoints are for resuming, not for reverting.
  • You work outside Git.

Related case

The full case for the project

FAQ

What people ask me most about AHP+

Does it commit or push for me?

No. It does not pull, merge, deploy or publish either. You authorize those.

Does my project have to use Node?

No. You need Node 20 or later and Git on the machine, and a repository with at least one commit. If package.json is missing, setup creates a minimal private one.

Which agents does it work with?

Codex, Claude Code, Cursor and OpenCode have an adapter, and ChatGPT a limited read-only one. Any other agent with access to the repository can follow the instructions in AGENTS.md.

What if the hand-off does not match the repository?

handoff receive compares project, commit, tree, branch and local changes. If something does not match, the result is RECONCILIATION_REQUIRED and the agent must stop until you sort it out.

Does it store my secrets or the whole conversation?

It does not copy chats or hidden reasoning, and device private keys live outside Git. But .ahp/ travels with your repository: do not keep tokens or .env files there.

Spec sheet

Key facts about AHP+

Version
1.4.1 (September 4, 2026)
License
Apache-2.0, open source
Requirements
Node 20 or later, Git and a repository with at least one commit
Dependencies
No third-party ones
Tested on
Ubuntu, macOS and Windows, with Node 20 and 22
Install
npx @jossuealcala/ahp-plus@1.4.1 setup .
Price
Free

Do you work with several coding agents?

Install it in a test repository and run ahp project check. If you want it adapted to your workflow or your team, write to me.

Talk about my workflow