The fragmentation of context across AI coding agents is a persistent workflow bottleneck. JoeBro, a new free, open-source Mac application, addresses this by installing a local MCP (Model Context Protocol) server that allows Claude Code, Codex, and OpenCode to share a single, persistent memory. Developed by user joexk1, the tool eliminates the redundant discovery process where each agent must independently relearn project specifics, such as database ports or test configurations, at the start of every session.

Architecture and Data Scoping

The core mechanism is a lightweight MCP server exposed via a SQLite file, requiring no external server or model calls for maintenance. The system scopes memory strictly by project, using normalized Git remotes or repository paths as unique identifiers to ensure consistency across machines. It also supports global and user-specific scopes, though the latter is explicitly excluded from the context provided to coding agents to prevent personal preferences from polluting technical decisions. All data remains local, with no account required and no information leaving the user’s machine.

Heuristic Maintenance Without LLM Overhead

To prevent memory bloat, JoeBro employs deterministic, arithmetic-based housekeeping rather than relying on expensive LLM calls for curation. The system uses Jaccard similarity for deduplication, reinforcing existing facts if token overlap exceeds 0.8, while treating numerical or identifier discrepancies as distinct entries. Facts decay over time with a 30-day half-life, starting at a strength of 0.5 and increasing to 1.0 upon recall. This ensures that actively used knowledge remains prominent, while stale information is archived after a nightly cleanup pass, capped at 150 active facts per project.

Provenance and Transparency

Trust in shared memory requires visibility. Each fact records its provenance, including the originating agent, model version, and machine, allowing users to trace the source of potentially erroneous context. The interface provides a transparent view of all stored knowledge, displaying strength bars, age, and kind (e.g., 'gotcha', 'decision', 'convention'), with options to edit, pin, or archive entries manually. A session-start brief injects up to 400 tokens of the strongest, most relevant facts into the agent’s context, ensuring immediate utility without overwhelming the token limit.

Key Takeaways

  • JoeBro unifies context for Claude Code, Codex, and OpenCode via a local SQLite-backed MCP server.
  • Memory maintenance uses deterministic heuristics (Jaccard similarity, decay rates) instead of LLM calls.
  • Facts are scoped by Git remote and include provenance metadata for auditability.
  • The tool is open-source (GPLv3) and installs with a single command on macOS 26+.

The Bottom Line

By rejecting the trend of server-side, LLM-managed memory in favor of local, heuristic-driven SQLite storage, JoeBro offers a pragmatic solution to agent amnesia that prioritizes speed, privacy, and user control. The project is available on GitHub at github.com/joexk1/JoeBro.