Persome

Install

Works with Claude Code, Codex, VS Code, Cursor, Windsurf, Claude Desktop. Read the repository before you give any tool access to your data.

npx agentmarket install connector/intuition-lab-personal-model

About

Use cases: agent-memory

No reviews yet. Installed it? Reviews come from verified installers only. How reviews work

Permissions

What this server can do and what it asks for, from its tool list and install config. Checked Oct 25, 2026.

Tools

10 read · 6 write · 0 delete · 6 not yet classified

Secrets

None asked for

Install scripts

Not scanned yet

Runs as

Python package

Tools (22)

Read from the source code; the running server may differ. Descriptions are the server’s own words; hidden characters are shown as [U+…].

• attention_trajectory

  Where the user's attention went, with DWELL (time actually spent). Answers "what did I spend my attention / focus / time on today / this morning / the last N hours". Prefer this over guessing from raw activity whenever the question is about focus or time spent. Built from the timeline's per-block attention locus (the code-resolved focused region of each window): contiguous runs of the same surface — a window, a terminal pane, a web page — coalesced with their dwell. Returns: - ``by_dwell``: surfaces ranked by TOTAL time spent, longest first (what actually mattered), each with ``dwell_…

• behavior_patterns

  **CALL FIRST for "who is this user / how do they work" context** — the learned behavior model, not raw memories. Returns the RESIDENT layer of the user's memory: a `root` apex (one synthesized "who is this person, what matters to them now" narrative) plus promoted behavior regularities (`faces` — patterns that survived two independent extraction signals AND resampling stability; each with evidence count + confidence), and evidence-backed observed workflows (`skills`) with their latest playbook. `rendered` is a ready-to-use text block for matching the user's working style without granting actu…

• correct_memory

  Update the user's memory when they tell you something in it is WRONG. Call this the moment the user corrects a belief about themselves — "Peach isn't my name, it's a colleague", "Research Team is an org, not a person", "Alex J. and Alex Jones are the same person", "I don't live in Beijing anymore". Pass their correction verbatim. This is a directed memory UPDATE (manage memory like model weights): it traces the wrong belief back to its source facts, supersedes them through the memory choke-point (receipts kept — reversible), or retypes/merges the entity, and logs the update. Downstream summar…

• current_context

  **ALWAYS CALL** for present-tense or ambiguous-pronoun questions about the user's state. Two high-value trigger patterns: 1. Present-tense: *"right now / currently / just now / what am I / what's open"* — this is the tool. 2. Pronoun with no in-conversation antecedent: *"that / this / it / the bug / the error / the file"* — the user is pointing at their screen, not at chat history. Never reply with "I don't have code/context to look at" or ask the user to paste something — call this tool first. If it comes back empty, then ask. Asking for a paste when this tool would have …

• entity_graph

  **CALL for "who is X / how does X relate to people & projects"** — the relation graph around one identity, optionally AS OF a past date. Returns the resolved canonical identity, its edges (predicate / label / evidence count / validity window), identities reachable within `depth` hops, and `chain_to_user` — how X connects back to the user. Pass `as_of="2026-03-01"` to ask about a PAST state ("who was his boss in March") — edges closed before or opened after that date are excluded. `include_shadow=true` additionally lists UNPROVEN extracted edges, clearly separated — treat them as hints, never …

• get_model_snapshot

  Return a bounded projection of the versioned Personal Model. The default ``section=overview`` returns model/build stats plus compact Root, Face, and Volume objects. Use ``points``, ``lines``, ``faces``, ``volumes``, ``root``, or ``receipts`` for bounded pages; pass the returned opaque ``cursor`` for the next page, or up to 20 exact ``ids`` for a focused selection. Aggregate evidence-reference arrays stay summarized unless ``include_evidence_refs=true`` on their paged section, and still cannot exceed the result byte budget. ``section=full`` returns a small CLI-export instruction because an unb…

• get_pending_model_work

  Return pending semantic session counts without invoking any model.

• get_schema

  Return the memory organization spec (file naming, what each prefix means). Rarely needed at query time. Useful only if you need to reason about WHERE a new fact would be stored, or explain to the user how their memory is organized. For normal "look up a fact" flows, use `search` / `list_memories` directly.

• list_memories

  **ALWAYS CALL FIRST** on the first personal-context turn of a conversation. List all memory files with descriptions + entry counts. Cheap (one SQLite query, no file reads), so the cost of calling is essentially zero. Call whenever the user asks about themselves, their schedule, preferences, or ongoing work — the response tells you which files exist and what they're about (e.g. `event-YYYY-MM-DD.md` for a given day's session-level activity log; `user-profile.md` for identity; `user-preferences.md` for habits; `project-*.md` / `person-*.md` / `org-*.md` for specific entities). If you're about…

• persome_setup_check

  Confirm that this model can call Persome memory tools.

• process_pending_model_work

  Process pending Personal Model sessions using this MCP client's model. This is the user-funded modeling path: Persome requests MCP Sampling from the originating client, so Codex, Claude, or another compatible agent uses its own authenticated model entitlement. Persome never receives the client's login token. The call is explicit and bounded because it may consume the user's agent allowance. Requires the client to advertise MCP Sampling with tools. Clients that only support ordinary MCP tool calls cannot use this path; configure an API/local provider or use a compatible agent instead.

• query_health_metrics

  Query owner-authorized wearable and health observations. Use for questions about imported steps, heart rate, resting heart rate, active energy, sleep stages, or workouts. Results are raw device observations, not medical conclusions. Filters are exact for `metric` and `provider`; time bounds are ISO 8601 and form [since, until).

• read_memory

  Read the full contents of ONE memory file the user has on disk. Use after `list_memories` / `search` points you at a promising file. Entries come back chronological. Supports `since` / `until` (ISO timestamps), `tags` (filter by any matching tag), and `tail_n` (most recent N entries only).

• read_receipt

  Dereference ONE memory receipt — the `⟨entry_id:path⟩` handles that `search`'s `chains` narrative and other tools hand out. Returns the full entry (content, tags, timestamps, validity window, confidence, `superseded` flag) plus `nearby_captures` — breadcrumbs to the NEXT evidence layer (follow with `read_recent_capture(at=…)` for the raw on-screen text or explicitly request its screenshot). Use when you need to verify what a chain hop or a fact is actually based on; this is the audit trail from any memory down to the moment on screen. A `superseded: true` entry is history, not the current bel…

• read_recent_capture

  Hydrate ONE raw screen capture — the actual visible_text, focused input value, URL, and (optionally) screenshot from the buffer. Use this whenever a compressed memory entry isn't specific enough (e.g. an event-daily entry says "edited main.py at 14:30" but you need the actual code, or "read article" but you need the text). Most event-daily sub_tasks include an inline `raw: read_recent_capture(at=..., app_name=...)` breadcrumb — call it verbatim. For keyword-driven searches across the whole buffer, prefer `search_captures` first; this tool fetches one specific moment. Arguments: at …

Signals

Stars

1,328

Forks

90

Open issues

7

Contributors

9

Weekly downloads

34

Last commit

Aug 11, 2026 (2 months ago)

Last release

Jul 12, 2026 (3 months ago)

Last package publish

Jul 12, 2026 (3 months ago)

From GitHub (fetched Oct 24, 2026) and the package registry.

Details

Registry name

io.github.Intuition-Lab/personal-model

Version

0.3.2

Auth

No auth

Packages

pypi personal-model

License

Apache-2.0

Repository

github.com/Intuition-Lab/personal-model

First seen

Oct 24, 2026

Last checked

Oct 24, 2026

Source

Official MCP Registrygithub-graphql