Meaning Model

The Meaning Model

Project website and getting started · Try the story viewer

Let an AI build an explicit world model, then reason, create and remember through it. The Meaning Model lets a language model turn its learned knowledge into processes, relationships and concepts that can be inspected, tested and revised. Each explicit structure can anchor further inference and abstraction, including categories the agent discovers as it works. Prior measurement is not required to propose an account; inferred explanations remain distinct from observations and can be tested against evidence.

Comparative judgments can be made explicit through a Cut: one declared unit is divided under a question among exclusive answers and an explicit remainder. Measurements retain their external units.

The Meaning Model's two linked architectural contributions are progressive resolution and a shared construction medium. Progressive resolution builds a world coarsely and opens the parts that matter; each opening preserves accepted commitments or revises them explicitly. In the shared medium, world records, concepts, interpretations and documents occupy one address space, so a reason or a sentence can address the exact record it concerns, and a revision can be made together with the accounts and passages that depend on it.

This repository brings together the theory, its compact grammar appendix, the Rust engine and MCP interface, and *The Book of Conditions*, a worked construction with a complete twelve-chapter manuscript.

See the changelog for release changes, upgrade notes, and unreleased work.

To use the tool, follow the package installation guide and connect it to your assistant. The assistant calls life_modeling_context for purpose-specific operational guidance. Reading the papers is optional.

Workflows over one model

One MCP server supports these workflows over the same engine and graph:

• Modeling. General-purpose, revisable world models of whatever you choose to model, such as a market, an institution or a technology, with an optional Jev estimator for cheap first estimates. See general-purpose modeling.
• Narration. The storytelling add-on writes fiction from a model, with an author model, whole-life character trends, scene review, alignment audits and deepening passes. In story chat the user plays a character in the story's world while the LLM plays everyone else, from what each of them could know at that moment, with or without a book. See the storytelling add-on guide.
• Human authorship with feedback. The human writes and decides; the LLM supplies feedback. Enable MEANING_MODEL_ADDONS=storytelling to use life_story_feedback, which prepares a read-only task from supplied text or an exact project revision. No existing world or fictional author persona is required for feedback on an excerpt. See the feedback guide.
• Agent and user memory. Model ongoing work and reported personal history, with preferences, goals, decisions, beliefs and learning linked across time. Capture useful records within the chosen scope and recover them across sessions when durable storage is configured. See the memory guide.
• Ideation. The alien add-on searches for solution mechanisms through invented worlds, following *Ontology of the Alien*, and curates them into a revisable map of idea families. See the alien add-on guide.

Modeling and memory are always available; the two add-ons are opt-in with MEANING_MODEL_ADDONS. Each workflow keeps what it does as revisable, inspectable structure. No workflow turns an estimate into an observation, a review into an accepted fact, or an idea into evidence that it works. All share macro-to-micro exploration, recursive discovery and attributed Understanding Nodes. Their purpose and the human's delegation determine what the agent may infer, record or change.

In memory work, the assistant begins with the project or life context and connects reports to the processes they describe. As new evidence arrives, it returns to the earlier interpretations and decisions that evidence may change. In writing feedback, the assistant reads the whole before diagnosing a passage, distinguishes what the text says from an interpretation or artistic alternative, and follows the consequences of the human's chosen revisions. Both workflows use the same method: the assistant starts with a coarse account, develops the parts that matter, and reconsiders the whole in light of what it finds.

Memory and feedback have their own life_modeling_context purposes: agent_memory for the agent's own work, user_memory for what a user reports, and human_author_feedback for writing feedback. For memory that survives a restart, configure a private LIFE_SIM_STATE_FILE; use one server process per database. A correction keeps the earlier account and is distinct from a real change over time. Dependent interpretations need review; they do not update automatically. Transcript capture is optional and off by default; life_memory_transcript_configure turns it on for one memory context. It stores only the visible messages that the caller or chat integration supplies, separately from the process model and its interpretations. The server never observes a conversation itself.

The construction record

Version 0.10.0 bundles a browser viewer. Ask the connected assistant to “Open this model”; life_model_viewer_open returns a local link to the chosen model or graph revision. No separate viewer checkout or special run folder is required. While writing, ask “Keep the viewer following as we work.” The assistant uses mode: "live" with a graph hash to follow saved revisions and retain your reading context. Exact read-only snapshots remain the default. The browser must run on the same computer as the MCP. See the viewer guide for supported views and access requirements.

Reading position is independent of world time: selecting a passage highlights its model links, while reading opens separately. Optional document spans stay attached to stable passage boundaries as text changes. Concepts, world events, author records and the telling can remain distinct while Understanding Nodes connect them freely; a connection does not make their clocks or authority identical.

The model and its Understanding Graph are the modeler's understanding, not a report about it: what is done and not recorded cannot be picked up by the next agent. Every Event that carries a Cut has a description of what happens in it, so its numbers mean something. Choices, ideas, predictions and voice decisions are recorded as Understanding Nodes linked to the events, Cuts and passages they concern, and outside reviews are recorded under their actual reviewers. life_model_outline shows the present state with its notes at a chosen depth, and life_construction_replay replays the whole development step by step, each note beside the records as they were when it was written. An agent continuing unfamiliar work recovers its history through replay and an outline before changing it. When the exact previously read revision and context are retained, the agent reads subsequent changes and the records relevant to the current task. life_construction_export and life_construction_import carry the whole history to another engine with the same hashes, so the worked examples can be replayed anywhere.

Choose the model your application needs

Model only what serves your purpose. Choose the processes, categories, relationships, and level of detail; use a supplied template, adapt its output, or author a model directly. A single process is a valid starting point. A browser, learning tool, or relationship application need not build the same account of a person, or a complete person model at all.

The intended use is an external model an AI can work with: connect observations across time, compare explanations, explore possible developments, and decide what to investigate next. The modeler supplies those interpretations; the tool preserves their structure and revisions and executes supported, declared laws. Categories and decompositions are part of what the modeler develops and tests, not a fixed inventory to fill in. Shared record and validation rules remain in force as the application vocabulary changes.

See application-specific categories for template choices and a runnable vocabulary-revision example.

General-purpose modeling and optional Jev estimation

Start with life_general_modeling_start. The general workflow supports domain-defined processes across technology, adoption, institutions, markets, physical systems and other applications. life_world_model_build constructs the initial model and graph from a compact scaffold. Its required context review starts with the enclosing system and longer-term developments before focal processes; represented context, unknowns and justified exclusions become Understanding Nodes. It also requires consideration of authored numerical judgments, useful native conceptual decomposition, and meanings across dates or perspectives. The agent may explain why a dimension is unnecessary or unresolved; the tool does not demand arbitrary scores or depth. These checks do not prove completeness or causal relevance. The builder's preview can evaluate Jev initialEstimate questions; applying that exact proposal adopts the initial values as estimates without another provider call. life_process_estimate uses bounded questions to create typed process-value proposals; life_process_estimation_record saves the exact proposal, process records and review in the graph. Estimates retain their status and evidence, and do not silently become observations in accepted runtime history.

Set MEANING_MODEL_ESTIMATOR=typesafe and TYPESAFE_API_KEY to use Jev for batched structured judgments. This is independent of the storytelling add-on and is off by default. The LLM chooses the scope and questions; tool code handles repetitive record construction. Reduced end-to-end cost and latency require measurement, including setup and review work.

Every fresh agent begins with life-sim://guide/start-here (what the tool is for and how to work in it) and the process breakdown, puts a rough model of the whole on the page, and reads life-sim://protocol/grammar, the operational protocol and its purpose guide before it writes records by hand or goes deeper, unless the user gives it other instructions with MEANING_MODEL_INSTRUCTIONS (see the MCP server guide). life_modeling_read serves these resources to clients that expose only tools. Read life-sim://guide/general-modeling for the complete workflow and limits.

To bring in a book, use life_document_import with UTF-8 text or a local text file. It automatically splits the source into ordered, linkable document nodes, preserving exact text and source positions. Imported text stays separate from accepted facts and the manuscript. See the source import guide.

Optional storytelling add-on

The authoring workflow below describes delegated story generation. For a human writing their own work, use human-author feedback instead; feedback alone does not create a project or rewrite the manuscript.

The narrative graph is the authoritative authoring record. Create the model and graph before developing story material; store candidates, seed draws and alternatives, drafts, assessments, selections, local revisions, and dis