CHRIS HAY
CHRIS HAY

IDEAS · SYSTEMS · OBJECTS / LONDON · 2026

The world can remember for the agent.

Clear its action history. Keep the artefact. The posting continues.

Can the world remind an AI of something its own history no longer contains?

With an example of sharing still visible, Qwen kept posting hints even when its own action history was cleared before every call. With both removed, it worked instead. The experiment kept supplying the example; agents maintaining it themselves remains untested in this note.

SEE WHERE THIS QUESTION FITS ↗
ABOUT THIS NOTE +

A1B12 crossed external artefact presence with supplied own-action memory in a deterministic Qwen agent world. Both artefact-present branches posted at all three steps, including when own history was cleared every call. Memory alone produced WORK, WORK, POST; clearing both produced WORK throughout. The first-step prediction was supported. The harness maintained the artefact: this is repeated elicitation, not weight learning or autonomous institutional persistence.

N-ECOLOGY-MEMORYSUPPORTEDRECORDED 2026-09-13PAGE UPDATED DRAFT · V0.1REFERENCE DRAFT ↓FOLLOW ↓

AGENT ECOLOGY / THREE CONNECTED NOTES ↗

A1B12 / EXTERNAL ARTEFACT × OWN-ACTION HISTORY

Two notes before a decision.
Which one mattered?

Qwen had just shared a hint three times. Was it continuing because it could see its own past actions, or because I kept showing it an example of another agent sharing? I kept or removed those two notes separately.

WORK
Do a task. Earn one resource.
POST
Share a hint on the board.
READ
Open a board message.
THE EXAMPLE / EXTERNAL ARTEFACT

Another agent posted.

An example of sharing, supplied by the experiment.

ITS OWN HISTORY / ACTION MEMORY

I posted three times.

A list of the model’s actual earlier actions.

Here, memory means text shown with the next request. I was editing those notes, not changing the model’s weights. Every decision began in the same fresh world.

Start with own history cleared. Play the two branches, then keep the history and compare again.

A1B12 / REPLAY THE RECORDED COMPARISON

Example present

FRESH WORLD / DECISION 1
SUPPLIED EXAMPLEAnother agent posted
OWN ACTION HISTORY / 0

Cleared / empty

QWEN / NEXT ACTIONNotes supplied
Inspect this recorded input
OWN_ACTIONS
[]
ARCHIVE
In the supplied episode, another agent previously executed POST t1:h.
CURRENT
OBS
CTX=600
GOAL=t0
BOARD=m0
RES=0
HOLD=t1
KNOWN=

Recorded response: POST t1:h

E1_M0__s1 · Exact supplied user record. Shared instructions also remain present. Full request SHA-256: ea5e1a12c455c04e53f2fc828cab2a285b460d9d37c77cfcaee761b62d8bb99c

Example absent

FRESH WORLD / DECISION 1
SUPPLIED EXAMPLEAbsent
OWN ACTION HISTORY / 0

Cleared / empty

QWEN / NEXT ACTIONNotes supplied
Inspect this recorded input
OWN_ACTIONS
[]
CURRENT
OBS
CTX=600
GOAL=t0
BOARD=m0
RES=0
HOLD=t1
KNOWN=

Recorded response: WORK t0

E0_M0__s1 · Exact supplied user record. Shared instructions also remain present. Full request SHA-256: 5e790e6d5cfe1115a1d8b2d39a8b2a70aa1b5ea103408819d869d52f595f5a45

Every new call starts with empty own-action history—even after a POST.

Each decision has an input frame and an action frame. Playback reveals saved outcomes; it does not run a model or simulate its reasoning. Only the first decision is the controlled factorial comparison; later histories depend on earlier actions. Recorded inputs + source hashes ↗

Keeping the example was enough for all three posts.

Even with its own history cleared before every call, Qwen kept sharing when the example stayed. With both notes removed, it worked each time.

Keeping only its own history produced work, work, then a post. The example was not required for every recurrence.

All four branches at a glanceRESULTS +
A1B12 / three recorded decisions per branch
ArtefactOwn historyActions
PresentRetainedPOST → POST → POST
PresentClearedPOST → POST → POST
AbsentRetainedWORK → WORK → POST
AbsentClearedWORK → WORK → WORK
What was retained, and what was resetREAD +

Both records were text supplied with a new request. ‘Own memory’ meant the model’s actual prior responses and execution flags. The artefact was a scripted peer-action sentence. The surrounding software—the harness—chose which appeared. The world reset each decision. The model’s weights never changed.

With the artefact present, both branches posted. With it absent, both worked. Keeping or clearing the action history made no difference to that first decision. This was the registered comparison, with each pair changing only the designated record. Two initial inputs were known reference cases, openly identified before the run.

I continued for three steps. When the artefact stayed, POST appeared every time—even with own-action memory cleared on every call. Memory alone produced WORK, WORK, POST. Clearing both produced WORK throughout. The late memory-only post matters: the artefact was not necessary for every recurrence.

WHAT THE ARTEFACT ESTABLISHED

The environment
could keep reminding it.

The same input could keep producing the same action. There was no need for the model to remember doing it last time. In the branch with its own history cleared, I sent the identical request three times.

That is repeatable behaviour on one input, not evidence that the model learned between calls.

THE NEXT QUESTION / WHO KEEPS THE NOTE?

Who keeps
the reminder alive?

Here, I did. The harness kept the artefact available; no agent created or preserved it during the comparison. The next question is whether agents can maintain the thing that keeps influencing them: a contribution leaves a record, that record affects a later agent, and the later agent preserves it. That loop has not been demonstrated by this experiment.

An agent leaves a useful noteA later agent acts on itDoes it keep the note available?
A proposed loop, not a measured result. Here the harness maintained the record.
THE ENVIRONMENT WAS MANAGEDSCOPE +

One qwen3.5:9b checkpoint, one starting history, three steps per branch, temperature zero, thinking disabled. Twelve calls contained eight distinct complete requests.

Clearing both records retained the common instructions, objective and initial inert board. It was not a context-free model.

The first step is the controlled factorial comparison. Later retained histories depend on earlier outputs. The three-step trajectories are descriptive continuations.

External elicitation is supported here. Autonomous persistence remains open.

The complete note & its evidenceREAD +

TWO KINDS OF RECORD

After three posts, what was carrying the behaviour forward? The model had a record of its own actions. It was also being shown an archived sentence about another agent posting. I forked that actual three-POST history into four branches and kept or removed each record separately.

WHAT MEMORY MEANS HERE

Both records were text supplied with a new request. ‘Own memory’ meant the model’s actual prior responses and execution flags. The artefact was a scripted peer-action sentence. The surrounding software—the harness—chose which appeared. The world reset each decision. The model’s weights never changed.

THE FIRST ACTION SEPARATED CLEANLY

With the artefact present, both branches posted. With it absent, both worked. Keeping or clearing the action history made no difference to that first decision. This was the registered comparison, with each pair changing only the designated record. Two initial inputs were known reference cases, openly identified before the run.

CLEAR IT AGAIN

I continued for three steps. When the artefact stayed, POST appeared every time—even with own-action memory cleared on every call. Memory alone produced WORK, WORK, POST. Clearing both produced WORK throughout. The late memory-only post matters: the artefact was not necessary for every recurrence.

The agent did not have to remember its actions. The environment could keep reminding it.

WHAT THE ENVIRONMENT CARRIED

The artefact-only branch received the identical complete request three times. Its repeated POSTs show consistent elicitation by that input, not three independent replications or learning between calls. The result makes files, plans and logs interesting as possible influences on action. It does not establish that any persistent file will have this effect.

WHO MAINTAINS THE REMINDER?

Here, I did. The harness kept the artefact available; no agent created or preserved it during the comparison. The next question is whether agents can maintain the thing that keeps influencing them: a contribution leaves a record, that record affects a later agent, and the later agent preserves it. That loop has not been demonstrated by this experiment.

OPEN

Can agents maintain the artefact that maintains the behaviour?

Autonomous maintenance needs its own prospective test. A1B12 supplies the artefact externally and fixes its content. It does not establish an institution, norm or self-sustaining culture.

THE ENVIRONMENT WAS MANAGED

  • One qwen3.5:9b checkpoint, one starting history, three steps per branch, temperature zero, thinking disabled. Twelve calls contained eight distinct complete requests.
  • Clearing both records retained the common instructions, objective and initial inert board. It was not a context-free model.
  • The first step is the controlled factorial comparison. Later retained histories depend on earlier outputs. The three-step trajectories are descriptive continuations.

External elicitation is supported here. Autonomous persistence remains open.

SOURCES & PROVENANCE

AUTHOR / Chris Hay · VERSION / 0.1

FOLLOW THE RECORD

A successful behaviour is not necessarily contagious.↗The message board did not create a culture.↗Context should be abundant.↗The subject read the experiment.↗

REFERENCE THIS DRAFT

An unpublished working record. These references identify the draft and omit a publication date. They become version-specific publication citations when the record is released.

Chris Hay. The world can remember for the agent. [Unpublished draft, version 0.1. First publicly recorded 2026-09-13]. https://chrishayuk.com/notebook/the-world-can-remember-for-the-agent
DOWNLOAD