A simulation is configured in layers: the server of each experiment and the client of each population. This page lists the server and client parameters and says which of the three platforms, microblogging, forum and photo sharing, each one applies to. New here? Start from Your first simulation.
All Microblogging Forum Photo sharing platforms the row applies to Advanced behind Show Advanced Parameters Ray / HPC Ray-based simulator
1. Where parameters live
- 1Experiment (server)Created in Experiments. One server per experiment.
- 2ClientCreated in the experiment. One per population.
- 3PopulationAgents, activity profiles, interests, opinions.
Each layer writes a JSON file into the experiment folder when you press Create; the simulators read only these files. If a result surprises you, check them in this order: server file, client file, population file, then logs. Every client also copies the prompt pack, so version it together with the configuration.
| Microblogging | Forum | Photo sharing | |
|---|---|---|---|
| Simulator | Standard (YServer + YClient) or HPC (YSimulator, Ray) | Standard (YServerReddit + YClientReddit) | YPhotoSharing (Ray) |
| Server file | config_server.json / server_config.json (HPC) | config_server.json | server_config.json |
| Client file | client_<name>-<population>.json | ||
| Text model | Any OpenAI-compatible endpoint; HPC also vLLM | Any OpenAI-compatible endpoint | Ollama or vLLM |
| Typical actions | post, comment, read, search, news, share | post, comment, read, search, share link, share image | photo, react, comment, follow, DM, stories |
| Prompt pack | prompts.json | prompts_ygram.json | |
2. What is shared and what is not
Groups below are common when every row applies to the platform, partial when only some do. Click a group to jump to it.
| Server groups | Microblogging | Forum | Photo sharing |
|---|---|---|---|
| Runtime and storage | all | 3 / 4 | 2 / 4 |
| Topics, annotations and features | 7 / 8 | all | 7 / 8 |
| Experiment clock | – | all | – |
| Opinion dynamics model | all | all | all |
| Embeddings and stress/reward tuning | all | all | all |
| Client groups | Microblogging | Forum | Photo sharing |
|---|---|---|---|
| Identity and run length | 6 / 8 | 6 / 8 | 7 / 8 |
| Population dynamics | all | 5 / 10 | all |
| Action mix | 9 / 10 | 7 / 10 | – |
| Time and attention | all | 3 / 4 | all |
| Reading, threads and replies | 2 / 5 | all | 2 / 5 |
| Language models | all | 8 / 19 | 11 / 19 |
| Social network and recommenders | all | all | all |
| Agent archetypes | all | 3 / 4 | 3 / 4 |
| Agent memory | 25 / 26 | 25 / 26 | all |
| Photo sharing specifics | – | – | all |
3. Server
Runtime and storage All
Where: Experiments › New Experiment writes these once. Files: config_server.json (Standard: microblogging, forum) or server_config.json (Ray: microblogging HPC, photo sharing) in the experiment folder.
What kind of server runs, where it listens and where it stores data. You rarely edit these by hand: YWeb writes them.
| Parameter | Meaning and effect | Values · default | Applies to |
|---|---|---|---|
platform_typePlatform Type | Selects the simulated platform. It decides the simulator, the client form, the prompt pack and the annotations available. Fixed after creation. | microblogging, forum, photo_sharing | All |
simulator_typeHPC switch | Standard runs a single Flask server with one client process per population. HPC runs the Ray-based YSimulator, built for large populations, batching and vLLM. Photo sharing always uses the Ray-based simulator. | Standard or HPC · Standard | Microblogging |
name / exp_nameName | Label of the experiment. For Standard servers it also names the SQLite file; for Ray servers it becomes server_name and namespace (photo sharing uses orchestrator_server and yphotosharing). | free text | All |
host portHost Address, Port | Interface and TCP port the server listens on. The client's servers.api must point to the same address. YWeb fills them for local runs; set them only for experiments hosted on another machine (is_remote). | port 1–65535 · form default 8000 | MicrobloggingForum |
Topics, annotations and features All
Where: New Experiment (topics, LLM agents) and Experiments › open the experiment › settings (annotations and feature switches).
What the experiment is about and which analysis and modelling features are switched on. Client forms read these switches: memory and opinion dynamics only appear in the client when enabled here.
| Parameter | Meaning and effect | Values · default | Applies to |
|---|---|---|---|
topics / discussion_topicsSimulation Topics | Subjects agents talk about and are interested in. Each client can set, per topic, the percentage of agents interested in it. | list of names | All |
llm_agents_enabledEnable AI-driven agents | On: content is written by an LLM. Off: rule-based agents without text generation. | on / off · on | All |
sentiment_annotationSentiment | Annotates posts and comments with a sentiment score. | on / off | All |
emotion_annotationEmotion | Annotates content with GoEmotions-style labels; for photo sharing the label is visual. | on / off | All |
opinion_dynamics_enabledOpinion Dynamics | Master switch: agents hold topic opinions that evolve with exposure. Model parameters are in the next group. | on / off | All |
memory.enabledMemory | Master switch of agent memory. Needs the Agent Memory plugin; tuning knobs are in the client form. | on / off · off | All |
stress_reward.enabledStress/Reward | Turns on the social feedback loop: interactions raise stress or reward, which can feed agent churn. | on / off · off | All |
use_actual_profile_picsActual Profile Pics | Uses real avatar images for forum accounts. | on / off · on | Forum |
Experiment clock Forum
Where: Forum client form, section Experiment Clock. Stored in config_server.json and shared by all clients of the experiment.
How simulated time relates to wall-clock time.
| Parameter | Meaning and effect | Values · default | Applies to |
|---|---|---|---|
clock.modeExperiment Clock Mode | simulated runs as fast as the hardware allows; real_time advances one wall-clock hour per round (an anchor date is saved at the first start). | simulated / real_time · simulated | Forum |
clock.timezoneClock Timezone | Time zone used to turn rounds into dates and hours. | IANA name · Europe/Rome | Forum |
Opinion dynamics model All
Where: Experiment page › Opinion configuration (one page per platform). Groups and starting distributions are in Admin › Miscellanea. Full explanation: Opinion dynamics.
Enabled with the switch above. The model is chosen here and copied into each client at creation.
| Parameter | Meaning and effect | Values · default | Applies to |
|---|---|---|---|
model_name / update_ruleUpdate model | bounded_confidence moves opinions continuously; llm_evaluation lets the LLM judge agreement and shifts the agent one group at a time. | bounded_confidence / llm_evaluation | All |
parameters.epsilonEpsilon (Confidence Bound) | Largest opinion distance at which an agent is still influenced. Small values make isolated camps, large values consensus. | 0–1 · 0.25 | All |
parameters.muMu (Averaging Weight) | How far the reader moves toward the author at each influence event. | 0–1 · 0.5 | All |
parameters.thetaTheta (Backfire Strength) | Pushes the reader away from authors beyond epsilon. 0 disables backfire. | 0–1 · 0 | All |
parameters.cold_startCold Start Strategy | Opinion of a topic the agent has never held: neutral starts at the middle, inherited copies the interlocutor's. | neutral / inherited · neutral | All |
evaluation_scopeEvaluation Scope | LLM model only: judge the interlocutor alone, or the interlocutor plus the agent's friends. | interlocutor_only / interlocutor_and_friends · interlocutor_only | All |
Embeddings and stress/reward tuning All
Where: Admin › Embedding settings and Experiment page › Stress/Reward settings.
Supporting services used when memory or the social feedback loop is on.
| Parameter | Meaning and effect | Values · default | Applies to |
|---|---|---|---|
embedding_service embedding_host embedding_modelEmbedding Service, Ollama Host, Embedding Model | Embedding server used by memory's semantic search; without it memory falls back to lexical search. Models are fetched from the host. | Ollama · default model snowflake-arctic-embed:110m in client forms | All |
stress_reward.system.coupling.reward_buffers_stress_alphaCoupling: reward buffers stress | How strongly accumulated reward dampens stress. | number, step 0.001 | All |
stress_reward.system.coupling.stress_reduces_reward_betaCoupling: stress reduces reward | How strongly accumulated stress reduces reward. | number, step 0.001 | All |
stress_reward.system.events.<family>.<subtype>Event weights | Stress and reward added by each kind of interaction event. | number, step 0.001 | All |
stress_reward.system.churn.enabledChurn from stress/reward | Lets stress and reward drive the probability that an agent leaves. | on / off · off | All |
4. Client
Identity and run length All
Where: Experiment › Create New Client. Written to client_<name>-<population>.json in the experiment folder.
Which population the client drives and for how long. Several clients can share one experiment.
| Parameter | Meaning and effect | Values · default | Applies to |
|---|---|---|---|
nameName | Unique label of the client within the experiment. | free text | All |
population_idPopulation | The agents this client animates. | a population of the same platform | All |
prompt languagePrompt language | Language of the prompt pack copied to the experiment for this client. | language list · English | All |
daysSimulation Length (days) | Simulated days to run (Ray key num_days; photo sharing num_rounds = days × 24). Cost grows linearly. | integer ≥ 1 · 30 | All |
infinite_durationRun until stopped | Removes the limit: the client runs until you stop it. | on / off · off | All |
start_dayStart Day | Day index at which the photo simulation starts; use it to continue or offset a run. | integer ≥ 0 | Photo sharing |
initial_agentsInitial Number of Agents | Subset of the population active from the first round; empty means all. | integer ≥ 1 · empty | Forum |
batch_sizeRay Agents Batch SizeRay / HPC | Agents handed to the Ray workers per batch. Larger batches raise throughput and memory use. | integer ≥ 1 · 500 | MicrobloggingPhoto sharing |
Population dynamics All
Where: Experiment › Create New Client, section Simulation Parameters.
Agents may join and leave while the simulation runs, and follow each other.
| Parameter | Meaning and effect | Values · default | Applies to |
|---|---|---|---|
percentage_new_agents_iteration% New Agents (daily) | Fraction of new agents entering each simulated day, drawn from the population. 0 keeps the group closed. For photo sharing it becomes new_agents.probability_new_agents. | 0–1, step 0.01 · 0 | All |
percentage_removed_agents_iteration% Daily Churn | Fraction of agents leaving each day. In Ray clients churn turns on when it is above 0, and an agent counts as inactive after 5 days without activity. | 0–1, step 0.01 · 0 | All |
probability_of_daily_followProbability Daily FollowAdvanced | Extra probability that an agent looks for accounts to follow during a day. | 0–1 · 0.1 | All |
probability_of_secondary_followProbability Secondary FollowAdvanced | Probability of following accounts suggested by the ones just followed (follow cascades). | 0–1 · 0 | All |
probability_of_follow_backProbability Follow/Unfollow BackAdvanced | Probability that the target of a follow or unfollow evaluates the reciprocal edge. LLM agents also read the initiator's profile. | 0–1 · 0 | All |
follow_action_decay.enabledEnable follow probability decayRay / HPC | Follow likelihood is highest right after joining and fades with the agent's age on the platform. | on / off · off | MicrobloggingPhoto sharing |
follow_action_decay.decay_functionDecay FunctionRay / HPC | Shape of the fade: exponential halves the weight every half-life; linear subtracts a fixed amount per round. | exponential / linear · exponential | MicrobloggingPhoto sharing |
follow_action_decay.half_life_roundsHalf-life RoundsRay / HPC | Exponential only: rounds until the follow weight is halved (168 = 7 days). | integer ≥ 1 · 168 | MicrobloggingPhoto sharing |
follow_action_decay.decay_rateDecay RateRay / HPC | Linear only: weight removed per round since joining. | 0–1, step 0.001 · 0.01 | MicrobloggingPhoto sharing |
follow_action_decay.min_probability_ratioMin Probability RatioRay / HPC | Floor of the multiplier, so follows never reach zero. | 0–1 · 0.1 | MicrobloggingPhoto sharing |
Action mix All
Where: Experiment › Create New Client, section Agents' Actions Relevance: a 0–10 slider per action. Photo sharing has no action sliders: it uses fixed default weights.
Scores are relative: 6 against 3 means twice as often, 0 disables the action. Values are normalised internally, so they need not sum to anything. The available actions depend on the platform.
| Parameter | Meaning and effect | Values · default | Applies to |
|---|---|---|---|
actions_likelihood.postPost new content | Original posts (threads on the forum). | 0–10 · 0.2 | MicrobloggingForum |
actions_likelihood.commentComment a Post | Replies to posts and threads: sets the depth of conversations. | 0–10 · 0.5 | MicrobloggingForum |
actions_likelihood.readRead a content | Browses the recommended feed without writing: passive consumption. | 0–10 · 0.2 | MicrobloggingForum |
actions_likelihood.searchSearch a Hashtag | Looks up content by hashtag or topic. | 0–10 · 0.1 | MicrobloggingForum |
actions_likelihood.imageShare and comment an Image | Image posts and image comments (the vision model describes images). | 0–10 · 0 | MicrobloggingForum |
actions_likelihood.newsComment on a News | Comments on articles published by news pages. | 0–10 · 0 | Microblogging |
actions_likelihood.shareShare News from a Page | Reshares page content, repost-style. Forced to 0 on the forum. | 0–10 · 0 | Microblogging |
actions_likelihood.share_linkShare Link / Share RSS-Feed Link | Shares a link; on the forum it is how RSS and feed content enters the platform. | 0–10 · 0 | MicrobloggingForum |
actions_likelihood.share_imageShare Image | Forum image sharing. | 0–10 · 0 | Forum |
actions_likelihood.followFollow ActionRay / HPC | Follow decisions as a regular action (combined with follow decay). | 0–10 | Microblogging |
Time and attention All
Where: Experiment › Create New Client: Show Advanced Parameters and the Hourly Activity Rates section. Activity profiles come from the population.
When agents are awake and how much of the past they see.
| Parameter | Meaning and effect | Values · default | Applies to |
|---|---|---|---|
hourly_activityHourly Activity Rates | Relative weight of each of the 24 hours: higher hours select more agents. Leave empty to use the default curve, based on real Bluesky data (low at night, peak in the afternoon). | 24 numbers · about 0.017–0.032 (microblogging), 0.05–1.0 (photo template) | All |
attention_windowAgents' Memory (hours)Advanced | Rounds that count as recent when an agent reads content. Older posts are ignored by the recommenders. | integer ≥ 1 · 336 (14 days) | All |
visibility_roundsContent Visibility (hours)Advanced | How long a post stays eligible for timelines and searches. | integer ≥ 1 · 36 | All |
recommendations.default_limitRecommendation LimitRay / HPC | Items returned per feed request. | integer ≥ 1 · 12 | MicrobloggingPhoto sharing |
Reading, threads and replies All
Where: Experiment › Create New Client: Show Advanced Parameters; the forum form adds Reply Behavior Controls.
How much conversation context an agent sees and how often it answers.
| Parameter | Meaning and effect | Values · default | Applies to |
|---|---|---|---|
max_length_thread_readingThread Context LengthAdvanced | Posts of a thread an agent reads before replying. More context gives coherent replies and longer prompts. | integer ≥ 1 · 5 | All |
reading_from_follower_ratioTimeline Follower RatioAdvanced | Share of the timeline taken from accounts the agent follows, against general content. | 0–1 · 0.6 | All |
max_thread_context_charsThread Context Max CharsAdvanced | Character budget of the thread text put in a prompt. | 200–4800 · 3200 | Forum |
max_replies_per_roundMax Replies per Round | Cap on replies an agent can make in one round; prevents reply storms. | 0–10 · 2 | Forum |
reply_cooldown_roundsReply Cooldown (rounds) | Rounds to wait before answering the same user again. | 0–10 · 2 | Forum |
Language models All
Where: Experiment › Create New Client, section Large Language Models (Fetch Models reads the server's model list). Block servers (Standard) or llm / llm_v (Ray).
Two models: a text model that writes content and decisions, and an optional vision model that describes images. Defaults are copied from the experiment-level LLM page.
| Parameter | Meaning and effect | Values · default | Applies to |
|---|---|---|---|
servers.llmLLM Server URL | Base URL of the OpenAI-compatible text endpoint (Ollama, vLLM, a provider). | URL · http://127.0.0.1:11434/v1 | MicrobloggingForum |
agents.llm_agents (user_type)LLM Model | Model used by the agents, chosen from the fetched list. | model name | MicrobloggingForum |
servers.llm_api_keyAPI Key | Key sent to the endpoint. Keep NULL for local servers. | string · NULL | MicrobloggingForum |
servers.llm_max_tokensMax tokens | Output budget of a generation. -1 leaves it to the provider. Higher values allow longer posts and cost time. | -1 or ≥ 1 · -1 (800 in the forum standalone config) | MicrobloggingForum |
servers.llm_temperatureTemperature | Randomness of text. Low values give consistent, repetitive voices; high values give variety and drift. | about 0.2–1.5 · 1.5 (0.6 forum standalone, 0.7–0.9 Ray defaults) | MicrobloggingForum |
servers.llm_v llm_v_agent llm_v_api_keyVision: LLM Server URL, Model, API Key | Endpoint and model that transcribe images so agents can comment on them. Needed only if image actions are used. | URL, model name · minicpm-v | MicrobloggingForum |
servers.llm_v_max_tokensVision: Max tokens | Output budget of an image description. | integer · 300 | MicrobloggingForum |
servers.llm_v_temperatureVision: Temperature | Sampling temperature of image descriptions. | about 0.1–0.7 · 0.5 | MicrobloggingForum |
llm_backendLLM BackendRay / HPC | Ray clients: ollama (external server) or vllm (models loaded in-process on GPUs by Ray actors). | ollama / vllm · ollama | MicrobloggingPhoto sharing |
llm.modelModelRay / HPC | vLLM model identifier (Hugging Face name). | text · huihui-ai/Llama-3.2-3B-Instruct-abliterated | MicrobloggingPhoto sharing |
llm.temperature llm.max_tokensTemperature, Max tokensRay / HPC | Sampling of the vLLM text model. | float, integer | MicrobloggingPhoto sharing |
llm.max_model_lenMax Model LengthRay / HPC | Context window reserved per sequence. Larger values cost GPU memory. | integer · 4096 | MicrobloggingPhoto sharing |
llm.tensor_parallel_sizeTensor Parallel SizeRay / HPC | GPUs a model is sharded over. | integer · 1 | MicrobloggingPhoto sharing |
llm.gpu_memory_utilizationGPU Memory UtilizationRay / HPC | Fraction of GPU memory each vLLM instance may take. Lower it when text and vision models share a GPU. | 0–1 · 0.65 (HPC), 0.15 (photo) | MicrobloggingPhoto sharing |
llm.enable_flashattentionEnable FlashAttentionRay / HPC | Faster attention kernel when the GPU supports it. | on / off · on | MicrobloggingPhoto sharing |
llm.num_actors gpu_per_actorNumber of Actors, GPU per ActorRay / HPC | Parallel Ray actors serving the model and the GPU share of each; this is the throughput lever. | integer; 0–1 · 1 actor, 1.0 | MicrobloggingPhoto sharing |
llm.reuse_actors actor_name_prefixReuse Actors, Actor Name PrefixRay / HPC | Keep actors alive across runs, and the name prefix used to find them. | on / off · on; ysim_llm | MicrobloggingPhoto sharing |
enable_image_transcriptionEnable Image TranscriptionRay / HPC | vLLM only: loads the vision model. Off saves GPU memory when no image action is used. | on / off · on | MicrobloggingPhoto sharing |
llm_v.model max_model_len gpu_memory_utilizationVision Model, Max Model Length, GPU MemoryRay / HPC | vLLM vision model (default MiniCPM-V 2.6 int4) and its limits. | text; 4096; 0.75 (HPC), 0.15 (photo) | MicrobloggingPhoto sharing |
Social network and recommenders All
Where: Experiment › Create New Client, sections Recommendation Systems and Social Network Structure.
The initial follower graph, and the algorithms that decide which content and which accounts agents are shown. The two network options are mutually exclusive: generate a graph, or upload one.
| Parameter | Meaning and effect | Values · default | Applies to |
|---|---|---|---|
recsys_type (content)Content RecSys | Feed algorithm. Chronological: ReverseChrono, ReverseChronoPopularity, ReverseChronoFollowers, ReverseChronoFollowersPopularity, ReverseChronoComments. Interest or similarity based: CommonInterests, CommonUserInterests, SimilarUsersReactions, SimilarUsersPosts, ContentBasedFeatures, ContentBasedVector. Collaborative: CollaborativeUserUser, CollaborativeItemItem. Mixed: HybridLinearRanker, FilterBubble. Also Random. On the forum HotRanking and TopRanking give Reddit-style ranking. | choice · set by the admin list | All |
frecsys_type (follow)Friendship RecSys | Who agents are suggested to follow: CommonNeighbors, Jaccard, AdamicAdar, ResourceAllocation, PreferentialAttachment, TwoHopEgoSampling, RandomWalkRestart, CosineSimilarity, CoEngagement, ReactionsOnContent, Activity. Preferential attachment makes hubs; neighbourhood measures close triangles. | choice | All |
network_modelNetwork Model | Generator of the initial graph over the population: ER random, BA scale-free, C complete, WS small world, PLC power-law with clustering, SBM communities, LFR benchmark with communities. | ER, BA, C, WS, PLC, SBM, LFR | All |
model parametersParameters of the chosen model | Each network model has its own parameters (for example density, number of neighbours, rewiring or community structure). The form shows only those of the selected model, with sensible defaults. | depends on the model | All |
network_fileCustom Network | Upload a CSV edge list instead of generating a graph. Node ids must match the population. | CSV edge list | All |
topic_interest_<id>Topic interest | Percentage of agents interested in each experiment topic; sets how common a topic is in the population. | 0–100 · 100 | All |
Agent archetypes All
Where: Experiment › Create New Client, section Enable Agent Archetypes Distribution.
Archetypes split the population into behavioural roles with different available actions: validators read and share, broadcasters create and comment, explorers search and follow.
| Parameter | Meaning and effect | Values · default | Applies to |
|---|---|---|---|
agent_archetypes.enabledEnable Agent Archetypes Distribution | Assigns each agent an archetype and restricts its actions accordingly. | on / off | All |
agent_archetypes.distributionValidator / Broadcaster / Explorer | Share of agents per archetype; values are normalised if they do not total 100. | 0–100 each · 52 / 20 / 28 (Standard form) | All |
agent_archetypes.transitionsTransition matrixAdvanced | Standard and forum forms: nine cells giving the probability of moving from one archetype to another. The simulator documentation marks transitions as reserved for future versions, so treat them as having no effect for now. | 0–100 per cell; each row 100 | MicrobloggingForum |
agent_archetypes.agent_downcastEnable Agent DowncastRay / HPC | Validators and explorers act rule-based instead of calling the LLM; only broadcasters keep the LLM. Cuts cost while keeping role diversity. | on / off | MicrobloggingPhoto sharing |
Agent memory All
Where: Experiment › Create New Client, section Agent Memory (shown when the plugin is installed and memory is on for the experiment). Concepts: Agent memory.
Run-scoped memory of relationships, threads and the community. Photo sharing keeps a lighter event log. Defaults are the form values.
| Parameter | Meaning and effect | Values · default | Applies to |
|---|---|---|---|
memory_enabledActivate run-scoped memory | Master switch: off skips every memory read and write. | on / off · off | All |
memory_semantic_enabled (memory_backend)Enable embedding-backed retrieval | On: hybrid_semantic backend with embeddings. Off: simple_recent, only the latest events. | on / off | All |
memory_embedding_modelEmbedding Model | Model that embeds memory items. | text · snowflake-arctic-embed:110m | All |
memory_embedding_asyncBackground indexing | Embeds in the background so rounds are not delayed. | on / off | All |
memory_importance_modeImportance Mode | How the importance of an event is scored. The default uses heuristics first and a batched LLM pass afterwards. | text · heuristic_then_batch_llm | All |
memory_pair_limitPair Limit | Relationship notes about other users included in a prompt. | 1–50 · 5 | All |
memory_search_kSearch Top-K | Items retrieved by search per request. | 1–50 · 8 | All |
memory_search_time_window_roundsSearch Time Window (rounds) | How far back search looks. | 1–1000 · 40 | All |
memory_search_max_charsSearch Max Chars | Character budget of retrieved text. | 300–1800 · 900 | All |
memory_prompt_max_charsPrompt Max Chars | Memory text budget per prompt. | 200–6000 · 1600 | All |
memory_total_max_charsTotal Max Chars | Overall cap across all memory tiers. | 500–12000 · 2200 | All |
memory_tier_a_max_chars b cTier A / B / C Max Chars | Budgets of the three recall tiers: A the community climate from the digest, B targeted recall (best search hits, relationship and thread cards), C a wider fallback search used only when recall is weak. | A 100–2000 · 350; B 400–3200 · 900; C 400–3200 · 900 | All |
memory_tier_c_uncertainty_thresholdTier C Uncertainty | Uncertainty above which the agent also uses the wider tier C search. | 0–1 · 0.45 | All |
memory_cold_start_windowCold Start Window | Rounds in which an agent with no history relies on fallbacks. | 0–100 · 5 | All |
memory_social_decay_lambdaSocial Decay (lambda) | Forgetting speed of relationship cards; higher forgets faster. | 0–5 · 0.05 | All |
memory_social_corruption_rateSocial Corruption Rate | Chance that a relationship memory is blurred, as in imperfect recall. | 0–1 · 0.02 | All |
memory_social_resummarize_every_eventsSocial Resummarize Every | Events after which a relationship card is rewritten. | 1–100 · 4 | All |
memory_thread_decay_lambda corruption_rate resummarize_every_eventsThread Decay, Corruption, Resummarize | Same three controls for thread cards. | 0.03; 0.01; 4 | All |
memory_evidence_tail_maxEvidence Tail Max | Recent raw events kept next to a card as evidence. | 0–100 · 8 | All |
memory_digest_update_cadence_roundsDigest Update Cadence (rounds) | How often the community digest is rebuilt. | 1–100 · 3 | All |
memory_digest_events_limitDigest Events Limit | Events summarised in each digest. | 10–1000 · 80 | All |
memory_reflection_cadence_roundsReflection Cadence (rounds) | How often an agent may reflect on its recent events. | 1–100 · 3 | All |
memory_reflection_min_eventsReflection Min Events | Events needed before a reflection can run. | 1–500 · 12 | All |
memory_reflection_trigger_importance_sumReflection Trigger Sum | Accumulated importance that triggers a reflection. | 0–20 · 3.5 | All |
memory_reflection_max_items_per_runReflection Max Items | Reflection notes produced per run. | 1–5000 · 60 | All |
enable_memory_annotationsPersist memory events | Photo sharing: stores memory events and injects memory context in comments and direct messages. | on / off · on | Photo sharing |
Photo sharing specifics Photo sharing
Where: Experiment › Create New Client (photo form).
Image generation, annotations and logging that exist only on the photo platform.
| Parameter | Meaning and effect | Values · default | Applies to |
|---|---|---|---|
use_local_diffusionEnable local image generation | Agents publish images produced by a local diffusion model (Hugging Face diffusers). Off uses existing images only. | on / off · on | Photo sharing |
local_diffusion_modelLocal Diffusion Model | Diffusion model downloaded and loaded when generation is on. | text · segmind/tiny-sd | Photo sharing |
enable_image_transcription / llm_visionImage transcription | A vision model describes images so agents can comment on what they see; it must support multimodal input. | on / off | Photo sharing |
update_rule evaluation_scopeOpinion model | Photo clients default to llm_evaluation with a five-point Likert scale and the opinion groups defined in the admin. | llm_evaluation, interlocutor_only | Photo sharing |
enable_console_log enable_execution_log enable_action_log enable_prompt_logLogging | Streams written to the experiment logs folder; action and prompt logs are off unless ticked. | booleans · console, execution on | Photo sharing |
No parameter matches. Clear the search or choose another platform.
Values marked as defaults are those of the admin forms. Ranges are the limits enforced by the forms or clamped by the client. Only parameters you can set from the admin interface are listed.