Abstract

Long-running social simulations need agents to remember the same world differently. Yet as shared context grows, agents that began with distinct personas can drift toward similar behavior. MindTrace addresses this problem by storing world events separately from each agent's memory of them. A personal memory retains how the agent learned an event, whether the information still holds, and how readily it can be recalled.

The architecture forms a causal loop: events create personal memories, recalled memories guide decisions, decisions create new events, and memories that actually influenced an action become easier to recall again. Recall combines relevance with current memory accessibility, while evidence strength determines whether an agent can take a high-consequence action or should verify first. An attention scheduler changes processing depth without changing the order of the loop. Mechanism-validation runs with 9 agents over 30, 120, and 240 simulation steps exhibit the complete loop, importance review, memory dormancy and return, context-triggered reactivation, and explicit judgments about memory state.

Keywords　long-term memory · social simulation · source-aware memory · retrieval · evidence-gated action · attention scheduling

---

Introduction

Why long-running agents converge

A long-horizon simulation is useful only if its agents develop distinct histories inside the same world. Personality descriptions and action templates are not enough: as shared context accumulates, behavior tends to repeat and agents begin to converge. A study spanning several models and forty simulated social days reports high rolling action repetition [1]. What keeps two agents different is not just what happened, but who learned it first, who reported it, whether it was verified, whether it later shaped a decision, and how long it has gone unused.

Consider a resident who hears from a neighbor that tomorrow's community meeting has been canceled. Another resident reads the official notice, while a third still remembers the original schedule. All three refer to the same meeting, but they do not hold the same evidence and should not make the same decision. A memory system that stores only the sentence “the meeting is canceled” erases those differences.

MindTrace treats memory as evolving internal state rather than accumulated text. Perception, recall, decision, action, and reuse form one loop, so an agent's history grows from what it experienced and what it did with that experience. The architecture asks three questions: how does an event become a source-aware personal memory, how much action can the available evidence support, and how does using a memory change future recall?

Contributions

1. Shared events, personal memories. The world keeps one record of what happened, while each agent retains its own account with source, time, and validity. Misunderstanding, hearsay, and correction remain visible instead of collapsing into one text.
2. Memory shaped by time and use. A memory becomes harder to recall when it is left unused and easier to recall when it guides a decision, is repeated, or returns through a related context.
3. Evidence-gated action. The evidence behind a memory is matched to the consequence of an action. When support is too weak, the next step is verification rather than an unsupported high-consequence action.
4. Attention-constrained processing. Activity, sleep, and the salience of new information determine whether a cognitive step runs in full, runs in reduced form, or is skipped.

Architecture Overview

MindTrace has three state domains: world, memory, and decision. Four transitions connect them into a loop. The world records what happened; each agent forms a personal memory of the event; retrieved memories support a decision; and the completed action returns to the world as a new event. A fourth transition updates only the memories that actually supported the action. The attention scheduler decides how deeply each step can run. Figure 1 shows the complete loop, and Table 1 states the responsibility of each transition.

In the meeting example, the cancellation notice is a world event. One resident sees it directly and another hears it through a neighbor, producing two memories with different source paths. If either resident changes a plan, that decision creates a new event. The memory used to justify the change is then reinforced, while merely retrieved alternatives are not.

Transition	Core question	Main product
World → Memory	How did this agent learn about the event?	A personal memory with source and time
Memory → Decision	What comes to mind, and is the evidence strong enough to act?	Retrieved memories and an action decision
Decision → World	How does a decision become a new fact observable by others?	A time-ordered world event
Decision → Memory	Which memories actually supported the action?	Updated accessibility and use history
Attention scheduling	How deeply should this moment be processed?	Full, reduced, or skipped processing

From Shared Events to Personal Memory

One event, different accounts

The world keeps events in the order in which they occurred, giving every agent the same temporal reference. Personal memory does not copy that event log. It records how the agent encountered an event: performing an action creates first-hand memory, seeing it happen creates witnessed memory, and hearing about it preserves the path from the speaker back to the originating event. Different accounts can therefore refer to the same underlying event without becoming identical (Figure 2).

Two identifiers keep this distinction intact. Event identity answers “is this still the same meeting?” and merges repeated exposure to the same event. Source identity answers “how do I know?” and separates direct experience, one person's report, and confirmation from an independent source. In memory research, this ability to remember where information came from is called source monitoring; in MindTrace, it determines whether information is treated as experience, report, or inference.

What was true may no longer be true

An agent may still remember that the meeting was canceled after it has been rescheduled. The memory remains part of the agent's history, but it should no longer justify a current action. MindTrace therefore stores historical content separately from current validity. Conclusions derived from several memories also retain links to their supporting evidence. When one supporting memory becomes invalid, the conclusion returns for review. This follows the persistent-node and dependency-propagation ideas used in truth-maintenance systems [2], [3].

How Memories Fade and Return

Accessibility as ease of recall

A memory should not remain equally available forever. MindTrace represents its current ease of recall as accessibility a. An active memory can appear through ordinary recall. A dormant memory needs a direct cue to return. An inaccessible memory has passed beyond retrieval in the current lifecycle. These states describe availability, not truth: an active memory can be wrong, and a dormant memory can still be historically accurate.

A fixed half-life would make every memory fade at the same rate. MindTrace instead lengthens the effective half-life of memories that matter or have been reinforced:

A memory is active when a ≥ 2⁻³, dormant when 2⁻⁵ ≤ a < 2⁻³, and inaccessible when a < 2⁻⁵. Because the decay function uses base two, those boundaries occur after three and five effective half-lives (Figure 3).

From Recall to Action

Two paths back to memory

Ordinary recall asks two questions: how relevant is this memory to the current situation, and how readily can the agent recall it now? MindTrace multiplies those two values:

A direct cue follows a second path. If the current situation identifies a specific event, MindTrace can return the older memory without placing it in the ordinary ranking. Memory research describes this as cue-triggered recollection [4]. The first path covers what comes to mind on its own; the second covers what returns when prompted. Once a memory has returned, related material can continue to activate through learned associations [5].

Evidence should match consequence

Recall decides what comes to mind; evidence strength decides what the agent may safely do with it. MindTrace evaluates four properties of the available material (Table 2) and places actions into three classes: reversible, verifiable, and difficult to reverse. The harder an action is to undo or check, the stronger its supporting evidence must be (Figure 4).

Component	Question answered	Effect on the action threshold
Source	Is this first-hand experience, direct observation, or a report?	First-hand material carries greater evidential weight
Independent sources	How many mutually independent sources raised the same event?	Separates social corroboration from repeated transmission
Current validity	Does the content still hold?	Outdated content remains in history but cannot support a high-consequence action
Source reliability	How has information from this source evolved in the past?	Adjusts the evidence class of reported material

When the evidence is too weak for a high-consequence action, MindTrace chooses a step that can reduce uncertainty: inspect the original notice, ask a participant, or wait for an independent source. In the meeting example, a resident who heard the news second-hand checks the official notice before canceling an important plan. That verification creates a new first-hand event, becomes a new memory, and may support the later action.

How Actions Change Memory

Every action returns to the world

Every completed action becomes a new world event and may enter another agent's next perception. MindTrace derives consequence from the structure of the action itself. An action with an explicit inverse is reversible. An action whose result can be checked through a new observation is verifiable. An action that changes another person's state and has no direct inverse is difficult to reverse. The evidence threshold therefore reflects what the action can do in the simulated world.

Only used memories are reinforced

Each decision records which memories actually supported it. Only those memories receive an update; appearing in a candidate list is not enough. The update increments u in Equation (3), recalculates the effective half-life, and restarts decay from the memory's current accessibility. In this way, use has a lasting effect on what the agent is likely to remember next.

Memory can also recover without directly supporting an action. Hearing the same event again increments repeated exposure d; hearing it from a genuinely independent source increments s; and a related context can reactivate it indirectly. Together, these paths let memories strengthen through use, return through new information, and still fade when nothing brings them back.

When supporting evidence changes

A conclusion can depend on several memories. MindTrace keeps those links, so a conclusion returns for review when one of its supporting memories becomes outdated or dormant. If current support is restored, the conclusion can become usable again. If a critical piece is missing, the conclusion remains part of the agent's history but cannot support a high-consequence action.

Attention as a Processing Budget

Attention budget

An agent should not process every moment at full depth. MindTrace treats attention as a limited daily budget. Waking time sets the available capacity, actions and incoming information consume it, and sleep restores it according to the agent's schedule.

Full, reduced, or skipped

Reduced processing preserves the minimum useful result: an agent may recall fewer memories or continue an existing plan instead of reconsidering everything. A skipped step records whether sleep, an ongoing action, or low attention caused the omission. Salience combines novelty, direct personal relevance, and conflict with what the agent already believes. The scheduler chooses only the level; each cognitive module decides what “reduced” means for its own work. This separation follows the module autonomy of blackboard scheduling systems [9]. Figure 5 shows one agent's attention over a simulated day — consumption, recovery, and the switching of processing levels.

Mechanism Validation

Experimental setup

We run continuous simulations with 9 agents, one simulated hour per step, at horizons of 30, 120, and 240 steps. These runs cover approximately 1.25, 5, and 10 simulated days. As the horizon grows, the short causal loop, importance review, and full memory-state cycle successively enter the observable range (Table 3), revealing how the mechanisms connect and operate together.

Horizon	Individuals × steps	Main observation	Representative readings
Short	9 × 30	All four transitions operate; verification creates first-hand evidence	5,288 world events; 405 use-based memory updates; 17 high-consequence evidence decisions
Medium	9 × 120	Importance review begins; attention produces mixed processing levels	367 importance reviews; 127 reduced segments
Long	9 × 240	Active–dormant–returned cycles appear; agents express judgments about their own memory	525 transitions to dormancy; 219 cue returns; 290 context-triggered reactivations; 116 memory-state expressions

What appears at each horizon

The 30-step run is long enough to expose the full evidence-to-action loop. Of 17 high-consequence actions that required evidence, 9 passed the evidence threshold and 8 became verification steps. Two of the 9 allowed actions used first-hand memories created by an earlier verification, making verification a source of evidence for later decisions.

The 120-step run reaches the review window of the default half-life, so importance can begin to change how quickly a memory fades. The 240-step run crosses three effective half-lives and exposes the full state cycle: 525 memories become dormant, 219 return through direct cues, and 290 return through related context. Of 944 importance judgments, 349 extend the memory's subsequent half-life. This order matches Equations (1)–(3) (Figure 6).

Validation results

Together, the three horizons show the complete MindTrace process: events form source-aware personal memories; relevance and accessibility bring memories into recall; the evidence threshold selects action or verification; and actual use reinforces what is more likely to return. As simulated time continues, importance review, dormancy, cue return, context return, and memory-state expression emerge in the designed order.

Related Work

Most long-term agent-memory research asks how memories should be written, compressed, and retrieved. MindTrace begins with the next question: once a memory has been retrieved, is its evidence strong enough to justify an action, and does using it change what the agent will recall later? Work showing that selective memory intervention can outperform continuous injection supports the decision to expose memory state only when it matters [6]. Human-inspired memory architectures similarly treat maturation and retrieval-time reconsolidation as central mechanisms [7].

The decay of accessibility, return through a cue, and activation of associated material connect MindTrace to classical work in memory psychology and cognitive modeling [4], [5]. Rechecking a conclusion when its support changes follows truth-maintenance systems [2], [3]. The attention scheduler borrows module autonomy from blackboard systems [9], while its distinction between deliberate processing and continuing an established behavior aligns with dual-process theory and habit research [8], [10].

Conclusion

MindTrace makes memory part of the causal state of a social simulation. The world records what happened; each agent remembers that event from its own position; accessibility determines what is likely to return; evidence strength limits what the agent can safely do; and actual use changes future recall. Attention adds a cost to this process by limiting how deeply an agent can think at each simulated moment.

The validation runs show the intended loop operating across several time scales. Verification can create first-hand evidence for a later action; important or reused memories fade more slowly; dormant memories can return through a cue or related context; and agents can express uncertainty about what they remember. MindTrace thus shows how individual differences can continue to grow from experience, information source, action choice, and memory use instead of remaining fixed in an initial persona.

· · ·

References