One person model.
Exact context
for each task.

Use the implemented evidence-backed memory core today. The owner-controlled read boundary remains a local preview while the runtimes are unified.

Agenttask + purpose
requests 4 fields
PersonOSowner decides
Project context granted Calendar window granted Health context not granted Private contacts not granted
short-lived capability
ContextPack 2 approved fields grant · provenance · receipt id

Memory without governance becomes surveillance. Permissions without a person model become forms.

Two loops. One boundary between them.

The core memory API is implemented in this repository. The owner-mediated authorization loop is a local macOS preview, not yet a public unified API.

Current core

Evidence-backed memory loop

Store the source first. Recall a neutral brief before answering. Consolidate later.

  1. register

    Issue a token for an isolated user memory space.

  2. recall

    Return a brief, memory atoms, and evidence before the agent answers.

  3. answer

    Your agent composes the response. PersonOS does not impersonate it.

  4. ingest

    Append the completed turn as evidence; extract memory asynchronously.

  5. light-dream

    Merge, reconcile, or preserve disagreement within the session.

  6. trace

    Walk from a memory to source evidence, or from evidence to cited memories.

Local preview

Owner-controlled context loop

Keep keys and the approval surface on the owner side. Compile only the granted slice.

  1. request

    Requester, process chain, fields, purpose, and duration travel together.

  2. partial grant

    The owner can approve fields one by one; high sensitivity requires step-up confirmation.

  3. ContextPack

    Denied fields return no preview or length signal. Approved fields receive a short-lived grant.

  4. receipt

    Read, deny, grant, revoke, and proposal events produce an auditable record.

  5. proposal

    An agent may suggest a durable update, but it cannot silently write owner truth.

One truth source. Multiple replaceable projections.

PersonOS keeps evidence durable and treats atoms, person models, indexes, and ContextPacks as rebuildable products of that evidence.

  • Evidence Immutable source record
  • Atoms Claim, fact, and event with time
  • Person model Stable, revisable context
  • Working memory Fresh session state
Evidence logthe source of truth
Memory atomsPerson modelWorking memory
Purpose + grantread boundary
ContextPackminimum visible context

Run the current core.

This quickstart reflects the local repository contract. It does not point to a hosted production service.

terminal
# from the PersonOS repository
python3.11 -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt

# configure model keys, then start
python run_dev.py

curl http://127.0.0.1:8000/api/v1/health
recall
curl -X POST http://127.0.0.1:8000/api/v1/recall \
  -H 'Content-Type: application/json' \
  -H 'X-User-Token: <token>' \
  -d '{
    "session_id": "demo-001",
    "query": "What context matters here?",
    "mode": "auto"
  }'
POST /users/registerCreate an isolated user space
POST /recallGet a neutral brief plus evidence
POST /ingestAppend a completed turn as evidence
POST /light-dreamConsolidate one session asynchronously
GET /trace/{id}Inspect the provenance chain
GET /tasks/{id}Read asynchronous task status

Where this changes the product, not just the database.

Continue across agents without exposing the whole person

Carry the active goal, decisions, constraints, and evidence needed for the next task. Keep unrelated private context outside the handoff.

Build a long-term assistant that can change its mind

Preserve what was believed, when it was valid, and which new evidence replaced it. Avoid turning one inference into permanent identity.

The release boundary.

A credible integration starts by naming what the current code does not yet guarantee.

Implemented

Text evidence and atoms, bitemporal state, fast/deep recall, session light-dream, user isolation, async ingest, recall, trace, and task APIs.

Prototype code

macOS owner approval, partial grants, L0–L3 sensitivity, short-lived capabilities, Touch ID step-up, signed receipts, and proposal-only writes.

Before production claims

Unify policy and memory runtimes; bind requester identity to signed processes; move keys to hardened storage; close inference and cumulative disclosure gaps; ship complete deletion and formal evaluation.

Let the agent ask.
Let the owner decide.

Start with the inspectable memory core. Keep the authorization boundary explicit while the preview matures.