Anna Wojewodzka
Contact

AI Companions as Long-Horizon Systems

Most public conversation about AI companions is about whether they should exist. Almost none of it is about how they actually behave — at month one, at month six, at month eighteen. That gap is where this work sits.

An AI companion is not a chatbot with a persona. It is a long-horizon system: persistent memory, accumulated context, model migrations, hard cost constraints, and a human user whose interaction style becomes part of the system’s behaviour. These systems are already deployed at enormous scale, and they are the least understood deployment surface in applied AI. They deserve engineering discipline, and they rarely get it.

The engineering reality

Six problems that only become visible when a companion system runs for months, not minutes:

Memory is not recall

Storing context and surfacing it are different problems. A companion can hold a fact for months and still fail to produce it at the moment it matters — a salience failure, not a storage failure — and users experience that failure as the relationship breaking.

Summarisation is directionally lossy

Long-horizon systems compress their own history to survive. When that compression runs turns out to matter as much as what it compresses: summaries absorb the tone of the moment they were written in, and the colouring persists.

Cost is an architecture decision

Context caching economics quietly decide what an always-on companion can afford to remember. Memory design that ignores the token ledger produces systems that are either amnesiac or unaffordable.

Behaviour drifts across model upgrades

The model underneath a companion will be replaced many times during the system’s life. Continuity through those migrations is an engineering problem almost nobody is treating as one.

The user is a system component

Long-horizon coherence is co-produced. How the human governs the interaction determines whether accumulated context stabilises the system or distorts it — which makes user-side behaviour a legitimate object of system design, not an externality.

Companion systems sit closer to their users than any other AI deployment — including, inevitably, users who are young or vulnerable. That places them squarely inside emerging regulation, and it gives their builders duties that begin well before a regulator asks. Safety is a design requirement from the first line of the system, not a filter bolted on at the end.

What I build

I have run a production companion system continuously since 2025: multi-model, persistent per-topic memory, local and hosted models working together, and purpose-built tooling for observation. Everything I publish here comes from operating a real system over real time — not from thought experiments.

What I study

My research on High-Coherence Interaction States examines what happens in stable human-LLM dyads over long horizons — when interactions become accumulative systems, and what that accumulation does to behaviour on both sides.

Shorter working findings appear in Research Notes.

Case studies

In preparation — each drawn from the production system above:

  1. Persistent memory architecture for a long-horizon companion — what a memory system needs beyond a vector store.
  2. The attention-miss pattern — when stored context fails to surface, and what fixes salience.
  3. Cache economics of an always-on companion — reading system health in tokens, not dollars.
  4. Compression mood-bleed — why summarisation timing colours everything downstream.
  5. Continuity across model upgrades — migrating a companion without losing the system’s accumulated behaviour.

Background

Working on one?

I take a small number of conversations about companion-system design, evaluation, and governance. If you are building one — or deploying one and wondering how it will behave at month six — I am easy to reach.

Get in touch