I gave Claude Code and KimiCode shared memory so they stop forgetting my project

Writing the body from only the facts you supplied — no invented versions, dates, or metrics.ConradLu2740 published **ProactiveAgent**, an open-source project that gives **Claude Code** and **Kimi Code** a shared long-term memory so preferences and project context survive across tools and sessions. The work is on GitHub at github.com/ConradLu2740/ProactiveAgent, with a Hacker News thread at item id 49193190 that had 1 point and 0 comments at listing time. The stated problem is simple: each coding agent starts cold, so builders re-teach the same stack choices, conventions, and corrections every time they switch hosts.

Mechanically, **ProactiveAgent** is a zero-runtime-dependency core plus an MCP server (`@proactive-agent/core` and `@proactive-agent/mcp`) that mounts over stdio. Agents call tools such as `memory_capture`, `memory_extract`, and `memory_recall` against a local store under `~/.proma-proactive/`, with project-scoped memory and optional explicit global share. Auto-extracted items default to pending until confirmed. A public verification story claims Claude Code wrote a memory and Kimi Code recalled it at full relevance with no extra config, and that correction and automation suggestions were accepted through `suggest_now` / `suggest_accept`. The package also exposes a local `/today` panel and optional Claude Code hooks for session start and end.

For engineers, the useful part is treating memory as a user-level asset instead of a per-tool cache. If you already bounce between **Claude Code**, **Kimi Code**, **Cline**, **Cursor**, or similar MCP hosts, shared capture and recall cut the tax of re-stating TypeScript rules, test gates, and project layout. The pending queue and confirm/reject path matter in real repos: wrong or hostile content does not silently enter the recall set.

In market terms this sits next to built-in agent memory and other single-host memory plugins, but the pitch is cross-host sharing plus restrained proactive suggestions rather than passive notes inside one IDE. Most competitors keep memory inside one product boundary; **ProactiveAgent** bets on MCP as the common bus and local files as the source of truth, with no cloud sync in the default design.

Practical next steps: install via `npm install @proactive-agent/mcp` and `npx proactive-mcp init` (Node 18+), or clone and run under Bun if you want to hack the packages. Watch whether project isolation holds under multi-repo work, whether the suggest budget stays quiet enough to trust, and whether more hosts beyond the Claude Code / Kimi Code / Proma verification set behave the same. The interactive story page on the project site is the fastest way to see the capture-to-recall path before wiring it into a production tree.