Hearmemory — Development skill for Claude Code

HearMemory

English | [简体中文](README.zh-CN.md)

**Shared memory for multi-agent coding: parallel subagents and agents handing work to each other (Codex, Claude Code, Cursor) share one project memory, and Jev, a small, fast judge model, checks every agent's claims against real evidence (test runs, diffs, commits), so the next agent knows what it can trust.**

Overview

HearMemory records what coding agents do in a project (commands, test results, file edits, commits) and what they claim ("fixed X", "tests pass"). A background worker asks Jev a few narrow questions about these records: do two mentions refer to the same code object, are two records the same event, does one claim restate another, and does the recorded evidence support or refute a claim. Every new session and every new subagent then receives a short brief with the findings of the other agents and the status of each claim, and every `git commit` is checked against the shared memory.

Features

• Multi-agent coordination. Parallel subagents see each other's results as they arrive. Sequential agents, including agents of different vendors, continue from where the previous one stopped.
• Evidence-checked claims. Each claim is labelled SUPPORTED, DISPUTED, REFUTED, SAME-SOURCE ONLY, and so on, based on independent test runs, diffs and commits, with full provenance (agent, session, subagent, commit).
• Fast, low-cost judge. Jev answers in about 0.5 s per call (benchmark median) and costs well under one cent per working session. Without an API key, HearMemory falls back to deterministic rules.
• Pre-commit check. Commits that rely on refuted claims, touch files with open issues, or follow a failing test run are flagged or blocked, depending on configuration.