Search past conversations
These MCP tools read the conversations that project memory keeps.
Search the messages a person typed
The MCP tool proxium_search_prompts reads the messages a person typed in the project's conversations, newest first, without tool results and system text. It takes query (words to find), conversation (one conversation), and since_days (1 to 90, 7 by default). Each page holds 50 prompts, and the cursor of the answer reads the next page.
The MCP tool proxium_conversation_edits lists the files that agents changed in the project's conversations, newest first. It reads the edit tool calls that Proxium keeps from each assistant message. The edit tools are edit, write, apply_patch, str_replace, str_replace_based_edit_tool, multi_edit and create_file, in any case. The path comes from the argument path, file_path, filePath or file. It takes since_hours (1 to 720, 24 by default), pattern (a part of the path), conversation and max_edits (1 to 500, 100 by default).
The MCP tool proxium_conversation_diff gives the diff of each file that one conversation changed. The old and the new text of its edit calls rebuild the file. A write gives the whole text, and an edit that follows it replaces a part. Each file holds added, removed, reverted, whole_file and a unified diff with 3 lines of context. It takes conversation, file (a part of the path), max_files (1 to 100, 20 by default) and max_diff_chars (1 to 20000, 4000 by default).
The MCP tool proxium_conversation_commits lists the commits and the landings of one conversation, newest first. It reads the id and the message from the result of git commit and jj commit. It reads the change id and the log entry from the result of desk change and desk land. A landing of desk also holds the page of its change on the desk server. It takes conversation and max_commits (1 to 200, 50 by default).
An agent that needs the words around a prompt reads the full conversation with proxium_read_conversation.
Find the conversations like one conversation
The MCP tool proxium_similar_conversations finds the conversations that are similar to one conversation. It takes conversation, the conversation to start from. The search reads the first message that a person typed in that conversation, with the wrappers removed. It then searches the conversations of the project with the words of that message. The conversation itself stays out of the answer.
A conversation that holds at least two of the words ranks above a conversation that holds one. A word in the title or the summary of a conversation counts also. A recent conversation ranks a little above an old one with the same words. A conversation that holds only one of the words stays out when the first message holds two or more.
Each answer has the title, the summary, the last turn and a snippet around the first matched word. It takes also subject (one person or one part of the work), scope, and max_conversations (1 to 50, 20 by default). A conversation with no message that a person typed gives the error no_prompt.
How a conversation ended
When the memory worker reads a conversation that has no new turn, it writes how the conversation ended. This step makes no model call. proxium_read_conversation, proxium_resume_memory and proxium_list_conversations show the outcome. proxium_search_memory adds it with include_outcome. MCP tools gives the rules.
Dump a project, and digest a week
proxium_project_context reads the project in one call. It gives the knowledge page and the live memory by kind. It also gives the newest decisions, the newest conversations, the files of the last 7 days, the cost and the newest relations.
proxium_weekly_digest gives one window, 1 to 30 days and 7 by default. For each day it counts the conversations that started, the conversations with a turn, and the memories and decisions added and closed. It also gives the files changed, the most common words of the messages a person typed, and the cost of the window. A day with no activity stays, with zeros. The day of a conversation is its start day, in UTC.
Both tools read GET /api/teams/{slug}/memory/activity. The route takes since_days (1 to 30, 7 by default), subject and scope, and reads a replica. A since_days outside the range answers 400.