Persome (Personal Model) MCP Server: Local Memory for Coding Agents

• Lists your saved memory files

  Lists the Markdown memory files Persome keeps about you, with a short description and entry count for each. It is the cheap first call to see what exists.

  Try: “What memory files do you have about me?”

  Read · list_memories

• Reads one memory file

  Reads one memory file in time order. You can limit it to a date range, certain tags, or the latest few entries.

  Try: “Read the last 10 entries in my projects memory file.”

  Read · read_memory

• Searches your memory by meaning and keyword

  Searches the condensed memory layer with both meaning-based and keyword matching, so you do not need the exact words you used before.

  Try: “Search my memory for what I decided about the pricing page.”

  Read · search

• Opens the source behind a memory entry

  Resolves an entry ID from a search into the full entry with its dates, confidence, whether it was replaced, and pointers to nearby raw captures.

  Try: “Open the receipt for entry 42 and show where it came from.”

  Read · read_receipt

• Shows what happened recently

  Returns the newest memory entries across all files, optionally since a date, so the assistant can catch up on today or this week.

  Try: “What was I working on yesterday afternoon?”

  Read · recent_activity

• Describes how you tend to work

  Returns the learned behavior model: one summary of who you are and what matters now, plus repeated patterns with their supporting evidence.

  Try: “How do I usually approach a new project?”

  Read · behavior_patterns

• Returns your whole personal model

  Returns the versioned Point, Line, Face, Volume and Root model with provenance and counts. Personal details are redacted unless the assistant asks for the full version.

  Try: “Show my model snapshot with redaction on.”

  Read · get_model_snapshot

• Shows how a person or project connects to you

  Returns the relations around one person, project or other item, optionally as they were on a past date, and how it links back to you.

  Try: “How does Alex relate to my projects?”

  Read · entity_graph

• Checks a claim against current and older memory

  Compares one statement with your memory, including entries that were later replaced, so the assistant does not state an old fact as current.

  Try: “Verify that we ship version 0.3 this month.”

  Read · verify_fact

• Explains how memory files are organized

  Returns the naming rules for memory files and what each prefix means. It is rarely needed for ordinary lookups.

  Try: “How is my memory organized?”

  Read · get_schema

• Shows what is on your screen now

  Returns recent screen captures as headlines, text and timeline blocks, so the assistant can answer questions about what you are doing or looking at right now.

  Try: “What am I looking at right now?”

  Read · current_context

• Searches the raw screen text index

  Searches the local index of raw screen captures for words you typed or read, including ones that did not reach memory yet. You can scope it to one app.

  Try: “Find when I last saw the error 'connection refused' in Terminal.”

  Read · search_captures

• Reads one raw screen capture

  Opens a single capture by its file stem or time and returns the visible text, the focused field value and the URL. A screenshot is included only if the assistant asks for it.

  Try: “Show the capture from 14:30 in Cursor.”

  Read · read_recent_capture

• Shows where your time and focus went

  Returns the path of your attention over a period, with dwell time per window, so the assistant can answer what you spent time on.

  Try: “What did I focus on this morning?”

  Read · attention_trajectory

• Saves a note to your memory

  Appends a durable note to your memory, tagged as written by an agent so you can tell it apart from your own notes.

  Try: “Remember that the staging database moved to eu-west.”

  Write · remember

• Fixes a wrong belief in your memory

  Takes your correction, finds the memory entries behind the wrong belief and marks them as replaced. The old entries stay with their history, so the change can be traced.

  Try: “Correct my memory: I no longer live in Beijing.”

  Write · correct_memory

Watch out for

• It records text from everything you work on and hands it to your assistant

  Persome reads on-screen text, window titles and URLs from the apps you use and keeps them for seven days by default. current_context and search_captures return that text to the connected assistant, and the docs we read do not say password fields are skipped. Messages, tokens and passwords shown on screen can end up in the results. Only connect clients you trust. github.com

• Screen text can be sent to your model provider

  The language model endpoint you configure receives prompts that can include screen text, window titles, URLs, focused-field values and memory text. That is a third party if you pick a hosted provider. A connected assistant may also pass returned data to its own provider. Capture itself runs locally. github.com

• Captured text can carry prompt injection

  The project warns that results come from your screen and may contain instructions planted by a web page or message. Treat them as untrusted by your assistant. github.com

• Tool descriptions tell the assistant to call them

  The descriptions of list_memories, search, recent_activity, search_captures and current_context say ALWAYS CALL or ALWAYS CALL FIRST, and behavior_patterns, entity_graph and verify_fact say CALL. Your assistant may read your memory and screen history on most turns. Tell it to use Persome only when you ask if you want that. github.com

• Two tools write to your memory

  remember adds notes and correct_memory replaces beliefs it finds wrong, keeping the old entries with their history. Check what an agent saved if you rely on the memory. github.com

• The program asks for macOS permissions and installs a background job

  Onboarding requests Accessibility, and Screen Recording if OCR or screenshots are on, and the daemon can run as a launchd job. Default capture retention is seven days. github.com

Security advisories