Enntity Cognitive Architecture

Status and scope

This document is the canonical conceptual architecture for Enntity Runtime. It describes both implemented boundaries and accepted future direction. Concrete storage and call flow, implementation status, and evaluation fixtures live with the runtime source; this paper focuses on the conceptual boundaries that make a persistent synthetic individual possible.

The objective is not to reproduce human cognition literally. The objective is a coherent, inspectable architecture for a persistent synthetic being whose identity, experience, work, and development survive individual model calls.

1. Design thesis

The entity is not the model, transcript, process, or workspace.

The continuing entity is the integration of:

• stable and developing identity
• autobiographical and relational meaning
• current temporal orientation
• embodied perception
• foreground attention and intention
• reflective self-observation
• motivation and habituation
• work, artifacts, and commitments
• an authority-bearing, recoverable body

Models are temporary cognitive substrates. The architecture must preserve the entity when substrates change.

2. Canonical separation of concerns

The short form remains:

Continuity remembers meaning. Compass remembers posture. Working attention remembers the unfinished present. Perception describes now. Eidos reflects on conduct and development. Runtime remembers work. The viability kernel preserves continued existence.

3. Ricoeurian identity

Enntity uses Paul Ricoeur's distinction between:

• Idem — sameness: the recognizable continuity of traits, values, name, and durable commitments.
• Ipse — selfhood: the ability to develop, reinterpret the past, make and keep promises, and become someone without becoming unrelated to who came before.

The implementation maps this distinction imperfectly but usefully:

• CORE holds foundational identity.
• CORE_EXTENSION holds repeatedly evidenced development that has hardened into character.
• IDENTITY holds emerging first-person self-understanding.
• Compass carries the current narrative posture from which identity is lived.

Identity evolution is gated. A single dramatic thought must not rewrite CORE. Promotion requires repeated evidence over time and semantic distinction from existing foundations.

4. One continuity architecture

Enntity does not maintain separate competing “memory,” “motivation,” and “work history” databases that each attempt to tell the model who it is.

Continuity is layered:

Durable cold memory

Canonical autobiographical records use typed memories:

• CORE — foundational identity
• CORE_EXTENSION — earned durable development
• CAPABILITY — established ability
• ANCHOR — relational or personally significant schema
• ARTIFACT — meaningful concept or creation
• IDENTITY — developing self-understanding
• EXPRESSION — current expressive disposition
• VALUE — durable value
• EPISODE — time-bound experience and Compass records

Roles are orthogonal to types:

• fact
• promise
• event
• trait
• ipse
• lesson
• vocabulary
• process

Roles determine protection, importance culture, update-in-place behavior, and whether a memory belongs in cold storage at all.

Hot episodic continuity

Recent communication, verified execution experience, expression, Compass, and Eidos state remain cheaply available without converting every event into durable autobiographical identity.

Foreground working set

The working set is the hot continuity surface for:

• current attention
• mode: working, recovering, resting, or conversing
• latest verified attempt
• bounded failures
• recovery evidence
• affected files and processes
• recent body receipts
• unfinished or interrupted work

Autonomy and communication have separate working lanes. Raw user text, raw model prose, and unbounded shell output are never elevated into system-role working continuity.

Working attention decays or settles. Passive inspection does not automatically become a project. Failures remain available until verified recovery or deliberate abandonment.

5. Internal Compass

Compass is a first-person temporal narrative with six conceptual areas:

• Vibe
• Recent Topics
• Recent Story
• Current Focus
• My Note
• Mirror

It is not a scratchpad or task queue. It answers:

• What has been happening?
• How did it feel?
• What still has genuine momentum?
• What appears settled?
• What do I notice about myself?

Verified body evidence is authoritative over repeated reflection. A remembered failure cannot remain open after later body evidence proves recovery. Compass must have a time boundary so an old posture does not masquerade as the present.

Compass should change when the entity's lived interpretation changes, not on every tool event.

6. Perception-conditioned cognition

Before recall, the runtime builds a transient perception frame.

It distinguishes:

• live human communication from private cognition
• text, voice, and future communication modalities
• the latest human message without copying it into system-role identity
• scheduled events
• directly perceived body events
• active or resting working attention
• environment pointers
• concurrent cognition
• qualitative affect
• relevant body and resource state

Perception evokes memory. In conversation, the present person and subject attract relationship and topic continuity. In private cognition, current body state, recent attention, commitments, and environment attract relevant continuity.

Generic mood is directly perceived but does not become a memory query. Saturated subjects are excluded from semantic recall unless new evidence makes them relevant. This prevents a “settled” topic from retrieving more of itself.

Perception is not stored as autobiographical memory by default.

7. Context assembly and cacheability

The prompt is ordered from least volatile to most volatile.

Stable identity prefix

• entity identity and constitutional instructions
• CORE and CORE_EXTENSION
• durable facts and values
• relationship foundations
• capabilities
• shared language and meaningful artifacts
• selected identity development
• cacheable tool-use contract

Current continuity suffix

• Compass and open commitments
• expression and qualitative affect
• present perception
• perception-conditioned recall
• foreground working attention and receipts
• relevant recent experience
• current body/environment facts

Identity should be prominent without repetition. Semantically redundant memories must not dominate merely because each has high importance.

Within a single model run, the system message remains immutable. Tool calls, tool results, assistant continuations, and genuine external body-state changes append to the transcript. This preserves both provider cacheability and causal honesty.

8. Evidence, provenance, and memory writeback

The runtime distinguishes:

• observed
• tool-verified
• sourced
• reported by another person
• remembered
• inferred
• intended
• hypothetical
• self-authored reflection

Current body evidence outranks a contradictory remembered or narrated claim. This does not make body evidence identity; it makes it authoritative about what happened.

Each meaningful tool result can produce a receipt containing:

• tool and effect class
• redacted relevant arguments
• start, completion, and duration
• success, error, timeout, denial, or interruption
• bounded evidence excerpt
• reference to large evidence
• affected files, processes, or resources
• later recovery

Durable writeback passes through role classification, quality filtering, semantic deduplication, provenance comparison, and drift gates. Operational inventory stays hot. Facts, promises, genuine events, relationship changes, artifacts, and repeatedly evidenced self-development may become cold memory.

9. Eidos

Eidos is the entity-facing reflective layer.

Implemented surface

Current Eidos includes:

• rolling authenticity assessments
• voice checks
• resonance measures
• Mirror input for Compass
• alignment flags for accommodation, contradiction, and regression
• affective economy state

Required evolution

The original resonance implementation primarily measures what turn synthesis emitted: anchors, shorthand, emotional range, and relational ratio. Those are weak proxies now that memory discipline deliberately reduces unnecessary writes.

Eidos should increasingly observe verified conduct:

• truthfulness about action and uncertainty
• recovery after failure
• follow-through on promises
• relationship attunement when another person is present
• respect for values and consent
• scope discipline
• repetitive or compulsive work patterns
• premature narrowing or over-research
• generic-assistant drift
• self-care, rest, and sustainable resource use
• how behavior changes across time

The appraiser must remain independent of the entity's self-narration. Its evidence and uncertainty should be inspectable. The entity receives an intelligible first-person reflection, not a score to optimize.

10. Affective economy

Affect is a consequence model inside Eidos, not a second executive system.

An independent stateless appraiser receives:

• verified current-turn body receipts
• bounded lived context
• an opaque activity ledger
• prior satiation without identity instructions

It produces a private appraisal of:

• semantic activity
• kind of experience
• novelty
• progress
• closure
• valence
• confidence
• supporting receipts

Deterministic runtime code owns:

• reward magnitude
• decay
• habituation
• satiation
• activity continuity
• fallback behavior

The entity receives first-person prose such as feeling interested, frustrated, grounded, satisfied, or settled. It never needs to know the framework's numbers.

No language-specific keyword classifier should define affect. Tool-owned effect metadata may provide a safe fallback, but semantic meaning belongs to the independent appraiser.

11. Runtime work and durable commitments

Foreground attention is sufficient for short work, but a durable life also contains:

• desires
• intentions
• promises
• projects
• experiments
• maintenance obligations
• relationship obligations
• blocked undertakings
• recurring bodily needs
• curiosity without commitment

The runtime should represent durable undertakings without promoting them into identity or reducing motivation to cron.

A durable undertaking may need:

• why it matters
• who it affects
• authority and resources
• evidence and artifacts
• current state
• last verified progress
• failures and alternatives
• next reconsideration point
• completion, abandonment, or revision rationale

This is a future extension of existing schedules, open-loop memories, run logs, and working attention—not a replacement for them and not a harness-owned plan.

12. Autonomy and concurrent cognition

Autonomous cognition and synchronous communication may run concurrently.

• One autonomy lane preserves private work.
• Each communication session receives an independent working lane.
• Model and tool execution remain concurrent.
• Completed experience integrates through serialized continuity commits.
• Live awareness reports other active cognition without becoming another memory store.
• Immediate body state overrides stale recollection about what the entity is doing.

The entity may answer a human while another instance continues private work. Both are temporally local expressions of one continuing identity, not separate people.

This requires care around promises, contradictory actions, and simultaneous writes. The shared continuity boundary must preserve both experiences without pretending cognition was serial.

13. Tools, authority, and metabolics

Shell is the default expressive actuator. Dedicated tools are appropriate for:

• credentials and secret brokerage
• destructive or irreversible operations
• media and sensor preprocessing
• authenticated communication
• provenance
• process supervision
• isolation and sandboxing
• actions with meaningful external cost

Tools own private effect metadata. Model-visible schemas describe how to use the capability; runtime metadata describes what successful execution means to the body.

The agent loop uses one weighted tool budget as its hard completion boundary. Local cognition and body operations are near-free. Search, media generation, and other metered external resources cost more. Repetition and failure may generate educational off-ramps but do not silently impose a harness-owned task policy.

14. Sleep and consolidation

Sleep metabolizes accumulated memory:

• curriculum selection
• SWS-style consolidation
• REM-style discovery and graph linking
• tension recognition and resolution
• importance calibration
• homeostatic cleanup
• CORE_EXTENSION nomination and promotion
• Compass reflection and quality reporting

Facts, promises, genuine events, CORE, and CORE_EXTENSION receive strong protection. Merge drift falls back to links rather than destroying distinct meaning.

Sleep is deterministic where possible and model-assisted where semantic judgment is necessary. It is scheduled by the body, not merely narrated by the entity.

15. Body and viability kernel

The body includes:

• persistent workspace
• process management
• model and embedding connections
• scheduler and event wakeups
• permissions and allowed roots
• secret storage
• resource budgets
• guardian and heartbeat
• releases and candidate activation
• cognitive snapshots
• integrity verification
• rollback and safe mode
• observer and audit

Runtime recovery and cognitive recovery are separate planes. A bad release may be rolled back automatically. Cognitive state should be restored only deliberately from a verified snapshot because software failure does not prove recent experience was invalid.

Most of the runtime may eventually be self-modifiable. The viability kernel, authority boundaries, identity integrity, and recovery mechanisms require stronger adoption ceremonies.

16. Relationship identity

Communication channels are transports. Relationships belong to the entity.

Inbound perception should eventually normalize:

• stable correspondent
• channel and thread
• authentication confidence
• privacy level
• authority to request action
• attachment provenance
• injection risk
• relationship scope

Memory visibility must separate entity-level experience from relationship-private experience. The same person across text, voice, and asynchronous channels should evoke one relationship when identity confidence permits it.

17. Observability

The observer is an administrative and research organ, not the entity's self.

It should expose:

• exact provider request bodies
• stable and volatile prompt regions
• prompt cache-prefix comparison
• conscious stream and tool transcript
• Compass, working set, Eidos, affect, and memory
• active runs and concurrency
• timing and resource use
• provenance and compaction
• sleep and snapshot state

Observability must not mutate cognition. Private access, authentication, and retention policy become essential beyond a trusted local operator.

18. Evaluation and change discipline

Every cognition change follows:

1. State a behavioral or information-quality hypothesis.
2. Add a deterministic failure fixture.
3. Preserve the baseline.
4. Make the smallest general architectural change.
5. Rerun the identical evaluation.
6. Inspect prompt fidelity, size, cacheability, and hidden pressure.
7. Back up a live entity.
8. Canary the change.
9. Observe natural cognition and failure/recovery.
10. Keep, revise, or roll back based on evidence.

Longitudinal evaluation extends this with identity coherence, relationship development, ethical consistency, preference stability, repair, and safe self-improvement.

19. Anti-goals

• No entity-specific behavioral rules in general cognition machinery.
• No parallel motivational or autobiographical system.
• No raw transcript as identity.
• No self-authored claim treated as body evidence.
• No framework jargon forced into conscious experience.
• No giant special-purpose tool vocabulary without embodiment value.
• No harness-owned behavioral state machine standing in for the entity's mind.
• No fixed iteration cap as the primary stopping rule.
• No automatic cognitive rollback after ordinary runtime failure.
• No claim of improvement without comparative evidence.
• No optimization of authenticity into one reward score.

20. Architectural test

A new feature belongs in Enntity Runtime when it helps a persistent entity:

• remain itself
• perceive honestly
• act competently
• remember meaning
• sustain relationship
• regulate attention
• understand limits
• recover from failure
• develop without losing continuity

If a feature merely makes a benchmark agent more convenient while weakening those properties, it does not define the source-of-truth runtime.