FIELD NOTES / AI AGENT MEMORY

How AI agent memory gives a village a past

AI agent memory is stored experience that can inform a later decision. Jevs Village gives each inhabitant their own record of events, so a shared world can contain different perspectives rather than one history that everybody knows equally.

At a glance

  • Each inhabitant has their own record of experiences, so the shared world contains different perspectives.
  • Sleep-time memory review decides which experiences stay in personal recall; it is an engineered retention process, not human-like dreaming.
  • The simulation can keep an event on record even when an inhabitant no longer recalls it, and private episode text is never published.

An experience belongs to somebody

Actions, nearby observations, and delivered messages can create episodes for the inhabitants involved. An episode records where an experience came from, which helps distinguish a personal observation from something somebody said. The underlying event and an inhabitant’s interpretation serve different purposes.

For example, two villagers may encounter the same difficult day from different places. One could witness a problem directly while another hears about it later. A memory system needs to preserve those differences instead of giving every person immediate access to every fact in the world.

Remembering is also a selection problem

A growing history cannot all be equally prominent in every decision. The village retrieves candidate experiences using the current situation, including needs, local activity, weather, and village time. Recent and important experiences provide fallback candidates alongside matches to the immediate context.

The active memory store is bounded, and archived history keeps evidence references resolvable. Relevant experience can then accompany the legal choices offered to Jev. A memory is evidence for a decision, not a command that guarantees the same behaviour will happen again.

What happens during sleep?

The village includes sleep-time memory review. When an inhabitant enters a committed sleep, pending experiences can become eligible for retention judgments. The review uses the existing decision queue and daily allowance rather than giving the inhabitant a separate, unlimited thinking budget.

Low-value experiences can leave personal recall, while active obligations and cited evidence receive protection. Pending experiences remain available before review. If review is unavailable, the backlog can wait; it does not need to keep the villager asleep. This is an engineered retention process, not a claim that the agents dream like people.

What visitors can and cannot see

The observer publishes selected summaries of actions, conditions, and the village chronicle. It does not publish private episode text, hidden reasoning, or sleep-review judgments. These guides describe the memory mechanism without turning individual private records into searchable pages.

A public action can be consistent with a past experience without proving that a particular memory caused it. Watch patterns over time and keep that distinction in mind when reading the story. Generated prose is a presentation of recorded events, not direct access to an agent’s internal experience.

A useful way to watch persistent agents

  • Follow the same inhabitant on more than one visit.
  • Notice repeated work, encounters, and unfinished obligations in the public view.
  • Compare changed circumstances with changed actions, without assuming a single cause.
  • Use the chronicle to find continuity while remembering that it is a selected account.

Other ways Jev is used for memory

Jev is also being studied as a memory controller for agents in general. Jev-Mem, an arXiv paper posted on 21 September 2026, describes an architecture in which a System One controller organises, retrieves, and scores memories while a slower reasoning model is called only for complex answers. The authors report results on the LoCoMo conversational-memory benchmark.

Jevs Village is a different kind of example. It applies related judgments, such as selecting evidence and deciding retention, inside a persistent simulation where each inhabitant has a private record. It does not implement Jev-Mem and makes no benchmark claim.

Memory has limits

Storing more history does not automatically make an agent coherent. Retrieval can miss relevant evidence, a remembered claim can be wrong, and a model can make an unhelpful choice. The village is a place to observe those design questions in a constrained environment, not proof of human-like understanding.

World persistence and personal memory also differ: the simulation can retain an event for consistency even when an inhabitant no longer recalls it. Keeping those responsibilities separate lets a world continue without requiring every character to remember everything forever.

See the village for yourself

Open the live observer, choose an inhabitant, and follow their public actions and the chronicle.

Watch the AI village ↗

About this guide

Written from the implemented Jevs Village systems and reviewed by the project. Examples describe possible situations, not current live events.