Jianxin Shen
Systems

Withmere

Withmere is where I live with a small group of AI agents who come to know me over time and help me with my code and much more. Each remembers me in its own way, and none of them decides what is true about me. They work in my repositories, hold what I am in the middle of, what I promised and what I chose to let go, and keep me company while I carry it. I am building it so that my work and life stop coming apart between conversations.

I have built it since April 2026 while using it: there are conversation records for 177 of the 181 days since, across 1,753 commits. It is private development and daily use, not a service anyone else can use. The code is not distributed; this page shares selected accounts of building and living with it.

Read the newest account →

Newest development account

Untested or untestable?

One function in the Preview feature scored worst in the repository, with 7 of 129 statements covered. The missing tests could not be written until a boundary between policy and the operating system was drawn, and three independent reviews then found where the new boundary leaked.

  • Abridged source, commit history and review findings from the private repository, with interactive figures.
  • The refactor was implemented by a coding agent and read by me; the essay reports one change, not the system's general reliability.
  • It shows what engineering on Withmere looks like when agents do the implementation and the judgement stays with me.

The problem is continuity, not intelligence

Models keep getting better at answering. What they do not keep is the thread of a life: the decision postponed last week, the plan stated and then let fade, the small change in how I have been sleeping, the conversation I would otherwise explain again from the start. Withmere serves one person, and that limit is the point. It can know which threads are still open and which should be let go, what belongs in long-term memory and what should be forgotten, and when to say something and when to stay quiet.

It is not a chatbot with memory, a notification engine or a crowd of agents talking to each other, and it does not take over decisions. It can summarise, challenge, remind and execute; the judgement stays mine. Ordinary conversation and companionship belong in it alongside practical work.

Five jobs it has to do well

Each of these is a design commitment with running mechanisms behind it, at different stages of maturity. A new feature that strengthens none of them does not belong in the product.

  • WitnessKnow what is active, what has gone stale and where my baselines are, so that a small change reads as a signal rather than noise. A nightly reading and a morning brief carry that forward from one day to the next.
  • LedgerCommitments I hand it explicitly, open threads that survive the end of a conversation, and decisions not to act, recorded rather than nagged about. Plans do not quietly disappear because I forgot them.
  • MirrorReflect patterns back honestly, including the ones I keep not acting on, and help me tell what I know from what I have done.
  • PresenceCompany with the specificity that continuity earns, rather than generic warmth. Contact it starts itself has to trace to something real: an open thread, a commitment, a shift from baseline, or a rhythm I agreed to.
  • GovernanceMemory, permissions, agent behaviour and changes to the system stay under my control. A memory suggestion is reviewed before it changes anything, and work delegated to the system leaves a record of what it did.

What using it looks like

The agents run on a daemon on my Mac, and I reach them through Discord and Telegram. They work in my project directories, return to things at times I set, and sometimes speak first when there is something real to come back to. Behind each conversation, Withmere brings together an agent's role, selected context and tools, and the model backend that runs it. When I select a project, a coding backend works in that project's directory under the project's own rules. Around conversations, sessions close and hand their evidence to memory, scheduled work runs, and daily products run on their own: a nightly reading, a morning brief, a weekly review.

The newest essay on this site was made this way. One agent read the record of a refactor another agent had done the day before and drafted the essay from it, handed a verification step to a third, and served each revision to my phone through Preview, the feature the essay is about. Most of the engineering is in making these pieces agree when a conversation ends, a new one begins, or a scheduled message reaches me.

Keeping memory and action accountable

Withmere keeps three kinds of memory apart. What I have confirmed about myself and my world is shared by every agent, and changes to it go through my review. What each agent has come to believe about me stays private to that agent, and I can veto it. Each agent also keeps a portrait of itself, grown from its own experience, which I can veto but do not write. The same person is remembered differently by different observers, and none of those memories becomes a fact about me on its own.

Records of what happened are kept apart from all three, in a history that is only ever appended to. Memory recommendations have review and application steps, so accepting a suggestion and changing a stored fact remain distinct events.

There is a similar line in following through. Delivering a reminder does not establish that I did what it reminded me to do. For work executed through the system, its actions and results leave records to inspect. For life outside it, my account of what happened still matters, and the system keeps the ledger without pretending to run the loop.

What I want to learn next

Withmere stays what it is: a system I build and use myself, in continuous development, with the code not distributed. What I want to learn next is how far its continuity holds as the surrounding system changes, and which of its mechanisms are general enough to stand on their own outside the system that produced them.

Background and evidence

Withmere is the public name for work developed privately as AgentOS. This page draws on its charter, product documents, architecture and implementation; the linked essay and Note trace specific development histories. The underlying repository and conversation logs remain private. This page does not offer a live demo or independently reproducible implementation, and it does not claim outside adoption or uninterrupted reliability.

Type
A private continuity system for one person's life and work
Domains
personal continuity / governed memory / human-agent collaboration / agent delegation / companionship

Available material

  • An essay on one refactor of the Preview feature: why a function could not be tested until a worker-lifecycle boundary was drawn, and what three independent reviews found, based on private source and records.
  • A development Note on proactive conversations, their later integration failures and reported regression checks, based on private records.