---
name: proxium-memory
description: >-
  Use the proxium project memory through the proxium MCP tools (proxium_*). Use at
  the start of any task in a project that has proxium memory, when the user says
  "continue", "what do we know about X", "did we decide Y", "remember this", or
  when you learn something a later session needs. This is the hub: it gives the
  loop and routes to proxium-memory-recall (read) and proxium-memory-save (write).
compatibility: Needs the proxium MCP server (https://proxium.tech/mcp) connected in the agent.
---

# proxium project memory

proxium keeps one memory for each project. Two sources fill it:

- **Observed memory.** The gateway reads the project's LLM calls and extracts facts,
  preferences, decisions, events and procedures from them.
- **Taught memory.** You and the team write memories through the MCP tools.

All tools are on the `proxium` MCP server. Your client can show them with a prefix, for
example `mcp__proxium__proxium_recall_memory`. Each tool takes an optional `project` (a
slug). If the person has more than one project, call `proxium_list_projects` first and
pass the slug.

If no `proxium_*` tool is available, the server is not connected. Tell the person to add
the MCP server `https://proxium.tech/mcp` to this agent (the steps for each client are at
https://proxium.tech/connect), and continue the task without memory. Do not guess what the
memory holds.

## The loop

1. **Before the task, read.** Call `proxium_recall_memory`. It returns the project's
   knowledge page, the most trusted memories and the newest conversation summaries.
   Follow the proxium-memory-recall skill.
2. **During the task, search** when you need one specific thing:
   `proxium_search_memory`, then `proxium_read_conversation` for the whole
   conversation, or `proxium_query_graph` for what the memory relates about one thing.
3. **After the task, write back.** Save what a later session needs, correct what was
   wrong, and rate the memories you used. Follow the proxium-memory-save skill.

## Trust

- A memory has an origin: `human`, `agent` or `observed`. A person's memory ranks
  first, and only a person can correct or remove it.
- An agent's memory waits in the console review list until a person confirms it.
- A memory that callers marked `wrong` two times leaves the recall until a person
  reviews it.

## Scope

Pass `subject` only for a memory about one end user of the project's own product
(their id, as the app sends it in `x-axon-memory-subject`). Leave it out for
knowledge about the project.

## Do not

- Do not save secrets, keys, passwords or personal data that the task does not need.
  The gateway screens every write, but the rule is yours too.
- Do not save small talk, or a fact that is true only for this conversation.
- Do not assert a fact from memory without the memory's text. Quote it.