RTFM moves Claude Code memory snapshots into the background
Version 0.55.0 makes the optional session-end hook quiet and single-pass, and fixes duplicate indexing of linked memory directories.
RTFM 0.55.0 changes its optional Claude Code memory hook so session endings no longer wait for a foreground snapshot pass. The maintainer says the hook now runs in the background, stays silent when RTFM’s Python interpreter is unavailable, and uses a lock so several sessions ending together start one pass. Existing users need to rerun rtfm memory --install-hook to install the new hook entry.
One snapshot pass, one directory entry
The release describes two connected problems: the global hook ran inside agent sandboxes where its interpreter could not be found, and each session end waited while it read memory across projects. The maintainer’s roughly five-second example is a local observation, not a general performance benchmark.
Version 0.55.0 also fixes a linked memory directory being indexed under two project names, which the notes say caused a SQLite uniqueness error. It indexes the directory once under the owning project and removes the duplicate entry. Another fix preserves accented text when writing Claude Code settings instead of rewriting it as Unicode escapes. The release and implementation commit document these changes; AITrending has not reproduced the reported failures.
Local retrieval with a broader memory scope
RTFM is an MIT-licensed retrieval layer, rather than a model or replacement coding agent. Its tagged English README offers a Claude Code marketplace plugin and a manual pip install rtfm-ai route for MCP clients such as Codex and Cursor. Python 3.10 or newer is required; the plugin lists Linux, macOS, Windows and WSL support.
The optional memory command snapshots Claude Code’s curated notes across ~/.claude/projects/*/memory/ into ~/.rtfm/memory.db. That is a machine-wide cross-project scope, distinct from a single project’s index. Installing the Claude Code plugin also changes local instructions and pre-grants its MCP tool permissions, so the integration has more reach than a read-only search binary. RTFM’s core retrieval needs no API key, while the agent and model using its results retain their own access requirements and costs.
