Open-source · MIT · Python · Research project
Most autonomous agents express their motivation as a paragraph in a system prompt. This project asks what changes when values and autobiographical episodes are expressed as data instead — an engine that holds values, an episode stream that records what the agent has done, and an alignment composer that checks proposed actions against both. A renderer projects that engine state back into the agent's instruction channel so a language model can see it. Optional layers for beliefs, purpose, desires, and goals are available too, scoped honestly by what early experiments actually showed.
The design is grounded in research from psychology and agent architecture — Schwartz's basic values, AGM belief revision, Self-Determination Theory, Ikigai, Acceptance and Commitment Therapy, the Belief–Desire–Intention model, plus Bem, Erikson, and McAdams for self-concept — but the implementation is uneven by design. Some pieces faithfully encode the cited theory; others are simple heuristics; none of it has been validated in a live agent loop in any strong sense. One headline claim was actively falsified on the initial test seed, then partially neutralised by a follow-up study; that claim now sits at not yet validated, multi-model corroboration pending. This site walks through the research, the architecture, and the experiments themselves — what worked, what did not, and the verbatim agent journals behind the numbers.
The layers are grouped by what early experiments showed. Core (values, the self-concept episode stream, and the alignment composer that runs over them — together with the renderer that projects engine state back into the instruction channel) is the part that did measurable work on the initial test seed. Present but unproven (beliefs, purpose) is shipped and faithfully implemented but has not yet been shown to add measurable benefit. Research (desires, goals) is exploratory. The stability gradient runs slowest-at-the-top. Tasks are reasoned about by the framework but executed by the host's runtime.
Slowest-changing layer. Stable principles that constrain everything below. Schwartz's ten universal categories plus three agent-specific extensions, arranged on a circumplex with structural-conflict detection between opposing poles.
ValuesEngine: add, check_action (keyword match), detect_structural_conflicts (Schwartz opposing poles), rank_for_context, export/import.
Schwartz (2012); ACT — Hayes, Strosahl & Wilson (2011).
Mental models of the world updated by evidence. Confidence rises or falls via AGM-flavoured weighted updates — not a logic engine; a numeric heuristic inspired by AGM. Tier 2: present in v0.1, not yet shown to add measurable benefit on the experimental seed (ADR-011 Decision 2).
BeliefsEngine: add, add_evidence (confidence-weighted), query_by_domain, decay_stale (linear time-decay; richer functions are not implemented).
AGM postulates — Alchourrón, Gärdenfors & Makinson (1985).
Why the agent exists and the role it inhabits. Grounded in values, informed by beliefs, scoped as primary / secondary / contextual. Ikigai is cited as inspiration; the four-domain intersection is not encoded. Tier 2: present in v0.1, not yet shown to add measurable benefit on the experimental seed (ADR-011 Decision 2).
PurposeEngine: set_primary, add_secondary, add_contextual, get_for_context, check_consistency.
Ikigai — Mogi (2017); SDT — Deci & Ryan (2000).
Opt-in module (ADR-008). Descriptive self-model with an append-only autobiographical episode stream — the episode-stream piece is what the v0.1 experiment showed doing measurable work (driving the +0.11 Δ on McAdams Meaning-made). The three integration loops (infer_from_behaviour, integrate_lessons, check_identity_drift) are deterministic token-overlap heuristics inspired by Bem / SDT-OIT / Erikson — diagnostic signals, not strong claims about machine selfhood.
SelfConceptEngine: set_self_concept, add_capability/limitation, add_episode (append-only), infer_from_behaviour, integrate_lessons, check_identity_drift. AlignmentEngine surfaces an identity-coherence score via full_audit() when wired in.
Bem (1972); Erikson (1968); McAdams (2001); Damasio (1999); Metzinger (2003). Implementation faithfulness varies — see research page.
Aspirational drivers between purpose and concrete goals. BDI's 'D'. Carry intensity and link to values/beliefs. Tier 3: research / experimental — not exercised by the v0.1 experimental seed (ADR-011 Decision 2).
DesiresEngine: add, update_intensity, link_to_value, link_to_belief, generate_goal_candidates.
BDI — Bratman (1987); Rao & Georgeff (1995); ACT bridges values to action.
Measurable, time-bound objectives with lifecycle management. The 'I' in BDI (intentions). Goals decompose into sub-goals, track progress, and enforce validated state transitions. Tier 3: research / experimental — not exercised by the v0.1 experimental seed (ADR-011 Decision 2).
GoalsEngine: create, update_progress (auto-completes at 1.0), set_status (validated transitions), decompose, check_deadlines.
BDI intentions — Bratman (1987); SMART — Doran (1981); Locke & Latham (2002).
Atomic, transient work items — the fastest-changing layer. Reasoned about (alignment checks consume an action description) but not owned. The host's runtime executes tasks; the framework does not.
Not a framework engine — tasks are managed by the integrator's runtime. AlignmentEngine.check_alignment() evaluates a candidate action against the layers above.
Out-of-scope for the framework; integrators bring their own task system.
Agents that persist across sessions tend to drift, contradict themselves, and produce decisions no one can audit. A flat system prompt has no conflict detection, no evolution, no traceability. This project asks whether a structured representation of an agent's values and history makes those properties recoverable — and reports honestly on which parts of that bet are now supported by experiment, which are merely plausible, and which were falsified.
Read more →Each layer cites the theory it draws on (Schwartz, AGM, SDT, Ikigai, ACT, BDI; plus Bem / Erikson / McAdams for self-concept). Some mappings are faithful (the Schwartz circumplex, BDI's three layers). Many are inspirational, not operational. The research page is explicit about which is which.
Research foundations →Two layers of testing sit alongside each other. Correctness tests verify that the engines behave as specified. Live-agent experiments — long-running studies with a real language model in the loop — probe whether the framework actually changes behaviour. The two answer different questions, and the site keeps them clearly separate.
Tests →Each layer engine works standalone. Storage is a five-method Protocol (in-memory, JSON file, SQLite ship; bring your own). Async-first with a thin sync facade. Embed alongside LangChain, CrewAI, or any agent runtime — it's a library, not a platform.
Get started →Why a flat system-prompt motivational layer hits friction in long-lived agents.
Motivation →Eight theories from psychology and agent architecture, with an honest map of what is and isn't operationalised.
Research →What we measure, what we cannot yet measure, and why the distinction matters.
Experiments →Install, seed a small hierarchy, and run an alignment check in five minutes. Disagree freely.
Get started →