Repository navigation
Describe when to use the memory tools instead of ordering the model to - #8
Merged
Merged
Conversation
Both the Claude and OpenAI directory policies reject tool descriptions that
push the model to call tools the user didn't intend ("AUTONOMOUSLY",
"never wait", "without being asked"). State the situations each tool is for,
keep the same triggers, and add that nothing is saved when the user asks not
to be remembered.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Merged
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Why
Both app directories we plan to list in review tool descriptions for wording that pushes the model to call tools on its own:
Our
instructionsand docstrings say "Follow this protocol AUTONOMOUSLY", "never wait for the user to ask", and "Call this proactively/immediately — without being asked". That is the most likely review rejection for both the Claude plugin and the hosted connector (the Go server reuses this text).Change
Wording only; the triggers are the same.
instructions: opens with why the tools are there ("The user connected EverOS so you can remember what matters to them"), then lists the situations for each tool. Drops AUTONOMOUSLY / never-wait / immediately.search_memory,add_memory,get_profiledocstrings: "Use this when …" instead of "Call this … without being asked".Not changed here:
record_trajectoryasking for the full task transcript is a separate review risk (OpenAI forbids requesting prior-turn arrays). That likely means leaving it out of the directory builds rather than rewording, so it's a follow-up.Check
ruff + 91 tests pass. Worth a quick manual run in Claude Code to confirm the model still saves a stated preference and searches on "like last time" without being told to.
🤖 Generated with Claude Code