Hippo 3.2 turns coding corrections into rules for the next Claude Code session

Hippo 3.2.0 aims to stop a familiar coding-agent loop: you correct the agent, start another session, and have to explain the same rule again. The November 2 release adds capture of qualifying follow-up instructions and a Claude Code session-start hook that puts saved user rules back in front of the agent.

Hippo is a separate memory tool, rather than a new model or a replacement for Claude Code. It stores memories in local SQLite with readable markdown mirrors and connects to coding tools through host integrations and an MCP server. The new release changes how it recognizes and presents a user’s instructions.

Corrections become instructions, rather than old notes

The release says a sentence in a later user turn can become a pinned rule when it has at least four words and either names code, such as a path, flag or function call, or constrains a choice with words such as “only,” “instead” or “before.” Its example is a preference to use pnpm instead of npm. This recognition step does not call a model.

Captured user rules now appear first under a dedicated rules heading, without the confidence labels and historical-note formatting used for ordinary memories. A new Claude Code SessionStart hook runs hippo session-rules to present the current project’s qualifying rules from its project and global stores. Existing users need to run hippo hook install claude-code again to add it.

The author reports a small replay in which agents followed the rules with the new session-start hook and failed to do so with the per-prompt hook alone. That is a project-reported experiment, not an independently reproduced reliability guarantee. Rules stored by older versions keep their previous tags and may need to be captured again before the new session-start path recognizes them.

Setup and data flow still matter

The MIT-licensed package requires Node.js 22.16 or later. Its documented route is a global npm installation followed by hippo init inside a project. Init can create the memory store, add blocks to existing agent instruction files, install user-level hooks, import git history and agent memories, and schedule a daily run. Flags let users leave out hooks, scheduling or learning. Installing the package alone does not enable every host integration.

The local store also needs a precise privacy description. Default BM25 recall uses no model or network call. However, the tagged README says hippo sleep sends memory text to Anthropic for extraction when an ANTHROPIC_API_KEY is set, unless extraction is disabled. Sleep can run at session end and in the daily job; optional API embedders and rerankers also call external services. Local storage therefore does not mean every configured operation stays offline.

Capture remains a sentence-shape heuristic. It can save something that was not meant as a rule or miss a correction without the recognized code or limiting words. The immediate benefit is a more direct route from an explicit correction to the next session’s instructions; it does not establish that the agent will always obey or that every preference will be captured.