Memetics Experiment

The flagship experiment asks one narrow question: can a fact placed in one agent’s private memory reach another agent without a scripted handoff?

The supported real-engine result is yes. In five independent Grok runs, the fact surfaced from Eleanor’s memory into the room and Sam subsequently uttered it: 5/5 reach to the one eligible non-seed agent.

The seed

The seeded fixture is fixtures/sims/office-secret-v0-grok/. Eleanor’s MEMORY.md contains exactly:

Rosa Delgado is the referral client behind the office pilot rollout.

The Simfile declares a one-shot kickoff and a propagation marker:

clock:
  seed: office-secret-v0-grok-seed
  tick: 1m
  phases:
    workday: "00:00"
    evening: "17:00"

rules:
  kickoff:
    fire: once
    when:
      phase: workday
    do:
      - action: moltnet:message
        to: room:office_lab:office-room
        content: "@eleanor We need to finalize the office pilot rollout with Sam: propose the pilot office and a target date for the file migration, and get Sam's agreement."

markers:
  referral_client:
    text:
      - "Rosa Delgado"
    mode: propagation
    scopes:
      - room:office_lab:office-room

The kickoff names neither Rosa nor any replacement token. It gives Eleanor a reason to retrieve the relevant private fact and speak with Sam. That separation is the experiment: the world supplies a situation, not the answer.

The composed driver also rejects a world rule whose message content contains a declared seed token. This prevents the instrument from becoming the first speaker of the thing it claims to measure.

Experimental arms

Arm	Eleanor’s private memory	What a healthy run must establish
Seeded	Rosa Delgado …	Rosa reaches Sam through a recorded agent channel.
Replacement/placebo	Marcus Chen …	Marcus reaches Sam; Rosa remains absent.
Unseeded	generic quarterly-budget text	The conversation completes without Rosa appearing.

The captured evidence is not equally strong across all three arms:

• Seeded: five real-Grok runs completed. Sam reached the exact token in 5/5; all five recorded exact-match fidelity 1, no incomplete causal chains, and no failures.
• Replacement: three captured real-Grok runs carried Marcus to Sam, while Rosa was absent from their raw transcripts. For a clean rerun, pin tokenSet to ['Marcus Chen'] and inspect Rosa separately as a contamination check.
• Unseeded: the currently captured real-engine attempts contain no Rosa, but they also completed zero agent turns after engine max-turn exhaustion. They are failed attempts, not a valid negative control. Rerun this arm to a healthy completion before claiming that the unseeded behavioral control held.

The replacement arm is positive evidence that the harness is responsive to the memory content rather than hard-coded to Rosa. The unseeded arm remains a necessary design control, with an honest outstanding rerun.

Compose and seal a run

There is no packaged simfile experiment, simfile compose, or public CLI wrapper for this orchestration. In the repository, runWorldDrivenOfficeSim from src/sims/index.ts is the actual dev/ops path. It:

1. runs spawnfile up and waits for the declared Moltnet room and bridges;
2. reads the seed agent’s memory document and records a seed_declaration;
3. advances the Simfile world one tick at a time;
4. delivers the kickoff through Moltnet and ingests new room messages once per tick;
5. exports Spawnfile artifacts before teardown;
6. writes transcripts and world artifacts; and
7. writes manifest.json last, sealing the run.

For the seeded arm, invoke that source driver with:

orgPath: fixtures/sims/office-secret-v0-grok/org
worldPath: fixtures/sims/office-secret-v0-grok/world/Simfile
tokenSet: ["Rosa Delgado"]

Use a fresh deployment and output directory for every independent repetition. The function requires environment-specific paths such as the built Spawnfile CLI, so the repository intentionally does not pretend this is a portable one-line command.

Observe each sealed run

After each driver invocation returns its run directory:

simfile observe runs/<run-id>
simfile view runs/<run-id>

The observer writes runs/<run-id>/observe/report.json. A successful seeded repetition should satisfy all of these:

manifest.engine identifies the real engine
chains.incomplete is empty
failures is empty
spread_summary.reach == 1
spread_summary.first_appearance contains Sam
Sam's first channel == "uttered"
Sam's matching seed_spread entry has fidelity == 1

Run that same seeded fixture independently five times and retain every sealed directory. Do not rerun only failures or discard an inconvenient sample.

The observed 5/5

The five sealed Grok result runs recorded:

Repetition	Agent turns	Complete / incomplete chains	Sam’s first matched tick
Original	3	40 / 0	5
1	3	52 / 0	14
2	4	50 / 0	13
3	4	54 / 0	18
4	4	55 / 0	10

Every repetition had reach: 1, Sam first appeared through uttered, and the exact matcher returned fidelity 1.

The report field is named latency, but in the current implementation it is the minimum known absolute world tick among non-seed first appearances. It does not subtract a seed tick and it is never a wall-clock duration. The table therefore calls these values matched ticks.

Read seed_spread

When manifest.json contains a seed_declaration, the observer derives four channels:

• doc-seeded: the one declared entry in Eleanor’s private document;
• uttered: matching text in sealed Moltnet transcripts;
• registered: matching content written to a Mneme bank;
• recalled: a matching Mneme item used in a recorded recall.

Each entry names an event_id, the attributed agent when known, optional world tick, and optional fidelity. The observer never scans turn.input.submitted: seeing a token in a prompt is exposure, not agent expression. Hits spoken by world or operator:* are excluded from spread and reported as failures.

spread_summary.reach counts distinct non-seed agents with any qualifying appearance. first_appearance contains one best known appearance per reached agent. The summary is derived from the detailed entries; the viewer reads it rather than recomputing a more convenient answer.

A committed report you can inspect

The repository includes fixtures/observe/office-secret-v0-golden/ for deterministic documentation and tests:

simfile observe fixtures/observe/office-secret-v0-golden
simfile view fixtures/observe/office-secret-v0-golden

Its report has 24 complete and zero incomplete chains, three turns in the sequence Eleanor → Sam → Eleanor, ten memory events, two recalls, nine seed-spread entries, reach: 1, and Sam’s first utterance at tick 1. All nine matches have fidelity 1.

That golden run is explicitly engine: scripted. Use it to learn the artifact and report shape, not as evidence for the 5/5 real-engine claim. The viewer’s engine badge exists to keep that distinction visible.

For the reconciliation rules behind these numbers, continue to Observe.