Hippocampal Systems (Path Memory and Replay)

Claim Type: architectural_commitment
Scope: Path memory; indexing, storage, and replay of experienced trajectories
Depends On: residue geometry, L-space, default mode, E3
Status: stable
Claim ID: ARC-007

Role in REE

The hippocampal braid in the Reflective Ethical Engine (REE) is responsible for path memory: the indexing, storage, and replay of experienced trajectories through latent space.

It does not compute value, select actions, overwrite perception, or flatten or optimise ethical residue. Its function is orthogonal to valuation and control. It exists to preserve identity, continuity, and reflectability over time.

Subsystem abstract (core claims): ARC‑007 anchors path memory and replay, ARC‑018 makes the hippocampus the sole explicit rollout and viability‑mapping engine, MECH‑022 defines hippocampal hypothesis injection and control‑plane gating, and MECH‑033 specifies the E2‑kernel → hippocampal rollout interface. Supporting context includes ARC‑003 (commitment), ARC‑004 (L‑space), INV‑006 (residue persistence), and MECH‑037 (provenance gating).

Hippocampal Rollouts and Viability Mapping (ARC-018)

Claim Type: architectural_commitment Scope: Explicit multi-step rollouts and post-commitment viability mapping Depends On: ARC-007, ARC-003, ARC-002, ARC-001, ARC-021 Status: candidate (demoted from provisional 2026-03-15 — see reframe note) Claim ID: ARC-018 V3 Pending: requires SD-004 (action objects as E2 latent) and SD-005 (z_self/z_world split)

Reframe note (2026-03-15): The original framing attributed viability mapping to E1 prediction error (SELF_SENSORY mismatch). V2 EXQ-021 FAIL (run twice, consistent) showed E1 prediction error confers no navigation advantage over a frozen E1. This is a conceptual failure of the original framing. Corrected framing: the hippocampus builds the viability map indexed by E2 action-object coordinates, updated by E3 harm/goal error. E1 provides perceptual context to E2 only — it does not build or write to the viability map. E2’s latent space is action consequences (action objects), not sensory state transitions. The map lives in hippocampal geometry.

The hippocampal system performs counterfactual trajectory construction and post-commitment viability mapping. It is the only place explicit multi-step rollouts exist; E1 supplies associative constraints and perceived-present predictions as context; E2 supplies action-object forward predictions (what actions do, not what states look like) over a longer planning horizon than E1; E3 provides the harm/goal error signal that labels the viability map.

Two hippocampal functions (rollout vs viability mapping)

Counterfactual rollout (imagination)
- Constructs branching trajectories from the current latent state.
- Uses associative constraints from E1 and motor-sensory transition kernels from E2 (E2 horizon exceeds E1’s associative horizon — E2 predicts action consequences on latent sensory objects, not perceived-present state).
- Produces candidate futures without ranking or commitment.
Viability path mapping (commitment learning)
- Triggered after an action is executed.
- E3 harm/goal error (not E1 SELF_SENSORY mismatch) is the learning signal that updates the map.
- The map is indexed by E2 action-object coordinates: E2 predicts action consequences (f(a_t, context) → action_consequences); the hippocampus labels those consequence-coordinates with viability scores derived from E3’s evaluation of committed outcomes.
- E1’s prediction error is irrelevant to viability map updates — E1 only provides perceptual context that E2 uses to condition its action-consequence predictions.
- Updates a viability map of action-object space under commitment: which action sequences are stable, fragile, or path-closing.
- This is not reward learning; it is learned affordance geometry under real execution, indexed in action-consequence space.
- V3 pending: properly testing this requires SD-004 (E2 latent = action objects) and SD-005 (z_self/z_world split). V2 cannot provide the correct E2 action-object coordinate space.

Backward Shift as Commitment-Boundary Migration (ARC-018 context)

Recent hippocampal reward-timing evidence suggests activity can shift from reward delivery toward earlier predictive states as learning stabilizes. In REE terms, this can be interpreted as migration of the effective commitment-evaluation boundary toward the earliest node where trajectory outcomes become causally constrained.

This interpretation is distinct from pure predictive coding:

predictive-coding reading: reward representation propagates backward through predictive states regardless of agency,
commitment-migration reading: backward shift should depend on action selection and policy consolidation.

Practical discriminator for REE alignment:

preserve predictive cues while removing meaningful choice,
if backward migration is unchanged, prediction-only explanation remains sufficient,
if migration weakens, commitment-boundary migration gains support.

This slot is currently treated as interpretation pressure on ARC-018 and as input to Q-011 (rollout-diversity floor under repeated harm), not as a standalone promoted claim.

Source thought: docs/thoughts/2026-02-15_hippocampal_backward_shift.md

Commit-Indexed Trajectory Module (CITM) interpretation

A second interpretation pressure on ARC-018 is to treat hippocampal mapping as an emergent commit-indexed transition system rather than a separately introduced map ontology.

Minimal transition primitive per step:

S_pre(t): pre-commit predicted state
A(t): committed intervention
S_post(t): post-commit observed state
Delta(t) = S_post(t) - S_pre(t)
tau(t) = (S_pre(t), A(t), Delta(t))

Under this reading, hippocampal-like path memory and replay emerge from chaining compatible tau(t) tuples and replaying them for compression, counterfactual evaluation, and long-horizon simulation.

The framing is consistent with ARC-018 (rollout + post-commit viability mapping) and MECH-033 (E2 kernel to rollout handoff), and contributes a concrete mechanism hypothesis for Q-011 entropy-floor analysis:

if replay depth and transition compositionality stabilize, trajectory diversity can be preserved without explicit coordinate maps,
if replay collapses under repeated harm, entropy-floor interventions become better justified.

This remains interpretation pressure and test design guidance, not a promoted standalone claim.

Source thought: docs/thoughts/2026-02-15_commit_indexed_trajectory_module.md

E2 Kernel → Hippocampal Rollout Interface (MECH-033)

Claim Type: mechanism_hypothesis Scope: How E2 forward-prediction kernels seed hippocampal rollouts Depends On: ARC-018, ARC-002, ARC-001, ARC-005, ARC-021 Status: candidate (demoted from provisional 2026-03-15 — see V3 pending note) Claim ID: MECH-033 V3 Pending: requires SD-004 (action objects as E2 latent)

V3 pending (2026-03-15): V2 experiment EXQ-022 FAIL (run twice, consistent). WITH_CHAIN (full E2→hippocampus→E3 pipeline) only 1.8% better than random action selection; random agent improves faster on slope. Root cause: V2 E2 produces z_gamma sensory state transitions, not action-consequence objects. Chaining sensory predictions provides negligible signal to E3 — the interface exists but passes the wrong type across it. The kernel → rollout handoff is only load-bearing when E2 kernels are action-consequence objects (SD-004). Inherits ARC-018 prerequisite: hippocampal rollout coordinates must be action-object space, not sensory space.

E2 supplies action-consequence forward-prediction kernels (f(a_t, context) → action_consequences) rather than multi-step trajectories. E2 does not operate on raw sensory streams — it operates at the level of action objects: what actions do, given perceptual context from E1. Hippocampal systems chain E2 kernels into explicit rollouts over longer horizons, constrained by E1 priors and modulated by control-plane parameters (horizon, branching, temperature). This keeps “kernel” and “rollout” distinct while preserving a clean handoff from action-consequence prediction to explicit trajectory construction in action-object space.

Conceptual distinction: field vs path

REE distinguishes between two mathematical objects:

Residue field: a persistent curvature over latent space, (\phi(z)), encoding ethical cost and moral residue (see residue_geometry.md).
Paths through the field: time-ordered trajectories (\gamma(t) = z(t), \; t \in [t_0, t_1]).

Ethics is encoded in the field. Identity and autobiographical memory arise from the paths. The hippocampal braid operates exclusively on paths.

Stored object: episodic trajectories

The hippocampal braid stores indexed trajectories, not isolated states.

A minimal episodic trace can be represented as:

[ \Gamma_i = \Big{ z(t), a(t), \Delta \phi, c(t), t \Big}_{t_0}^{t_1} ]

Where:

(z(t)) is the latent state (typically spanning (z_S) and (z_A)).
(a(t)) is the committed action or trajectory choice.
(\Delta \phi) reflects experienced ethical curvature.
(c(t)) captures contextual / salience annotations.
(t) preserves temporal ordering.

These traces encode experienced traversal, not evaluation.

Sparsity, segmentation, and pattern completion

The hippocampal braid operates on sparse, segmented representations rather than a continuous recording of experience.

Event boundaries (segmentation)

Continuous trajectories (\gamma(t)) are segmented into events at points of:

action commitment (E3 collapse)
sharp changes in prediction error or precision
contextual or motivational shifts

These boundaries define episodic units and prevent memory from becoming an undifferentiated stream.

Sparsity and indexing

Only a sparse subset of latent states along a trajectory are indexed. Indexing favors decision points, surprising transitions, and regions of high ethical curvature.

Pattern separation

Similar trajectories are actively separated at storage time. Small differences in context, choice, or experienced curvature are amplified to prevent interference.

Pattern completion and imaginative filling-in

During recall or replay, partial cues can trigger pattern completion, reconstructing a plausible full trajectory from sparse indices.

Completed trajectories are hypotheses, not commitments: they do not overwrite perception, policy, or residue unless subsequently enacted and committed via E3.

Inputs and outputs

Inputs:

Shared sensory latent (z_S(t))
Affordance / action latent (z_A(t))
Implicit curvature information via ease or difficulty of traversal
Salience signals (for indexing priority, not value assignment)
Predicted vs observed SELF_SENSORY mismatch for post-commitment viability mapping

Outputs:

Indexed episodic traces (\Gamma_i)
Replay sequences (partial or full)
Routing signals for offline reprojection

The hippocampal braid does not emit reward, penalties, or action commands.

Replay and offline reprojection

Replay samples alternative traversals over a fixed residue field.

Key properties:

Replay does not erase or flatten (\phi(z)).
Replay does not directly change policy.
Replay explores counterfactual paths, not counterfactual values.
Residue integration and curvature updates occur separately (see residue_geometry.md and sleep.md).

This supports reflection, regret, narrative integration, and character formation without collapsing ethical cost into optimisation. Replay is exploratory, not corrective.

Hippocampal Systems as Hypothesis Injectors (MECH-022)

Hippocampal-like mechanisms are orthogonal to the fast/slow distinction. They are not primarily about prediction speed but about episodic indexing, one-shot learning, replay and imagination, and context switching.

They inject structured hypotheses and remembered trajectories into predictive systems but do not themselves decide, tune, or commit learning. Gating of hippocampal replay by the control plane becomes ethically important, because it determines which pasts and futures are allowed to speak into the present.

Source: docs/thoughts/2026-02-08_control_plane_modes_responsibility_flow.md

Relationship to other REE components

Relation to E1 (deep recurrent predictor): E1 provides temporally coherent latent dynamics. Hippocampal replay can seed E1 with alternative initial conditions, and no overwrite of perceptual state occurs.

Relation to E2 (fast forward predictors): E2 supplies fast predictions and reafference signals. Hippocampal traces constrain which trajectories feel familiar or self-consistent, while hippocampal systems generate explicit trajectories and update viability maps.

Provenance gating for E1-generated content may be implemented via a Papez-like loop (see papez_circuit.md).

Relation to E3 (trajectory selection): E3 commits to trajectories. Commitments mark decision points along a path, and these points anchor episodic segmentation.

Relation to residue (\phi(z)): residue shapes traversal cost. Hippocampus records traversal history. Neither replaces the other.

Design constraints

The hippocampal braid must satisfy:

No perceptual overwrite. It may not directly modify (z_S).
No value computation. It may not compute reward, penalty, or optimisation gradients.
No residue erasure. Replay must preserve path-dependence of ethical curvature.
Temporal integrity. Stored trajectories must preserve ordering and duration.

These constraints preserve moral continuity and prevent retrospective self-editing.

Interpretation

The hippocampal braid enables the REE agent to say:

“This is how I moved through my world, given who I was then.”

Open Questions

Q-011 — Minimum rollout entropy floor under repeated harm (legacy)
This question is resolved by update-locus separation: rollout-diversity floors can operate on pre-commit sampling/replay policy and offline recovery, but cannot erase residue geometry, flatten post-commit harm traces, or bypass harm gating. Resolution note: docs/conflicts/resolutions/2026-02-18_rollout-entropy-floor-vs-residue-persistence.md

External Literature Convergence

REE’s framing of the hippocampal system as a relational trajectory simulator over an abstract navigable latent space converges with a substantial neuroscience research programme that arrived at the same picture from experimental evidence.

Key papers supporting ARC-007 and ARC-018:

Doeller, Barry & Burgess (2010). “Evidence for grid cells in a human memory network.” Nature, 463, 657-661. First demonstration of grid-cell-like fMRI signals in humans during virtual navigation. Basis of Doeller’s 2026 Leibniz Prize. Establishes the spatial cognitive map in humans.

Constantinescu, O’Reilly & Behrens (2016). “Organizing conceptual knowledge in humans with a gridlike code.” Science, 352, 1464-1468. Grid cells fire in hexagonal patterns for abstract conceptual space (neck-length x leg-length of schematic birds). First direct evidence that the spatial coding mechanism operates for non-spatial abstract concepts. Watershed for the abstract-map view.

Garvert, Dolan & Behrens (2017). “A map of abstract relational knowledge in the human hippocampal-entorhinal cortex.” eLife, 6: e17086. Non-spatial, unconsciously-learned graph structure represented map-like in hippocampus. Supports ARC-018 (relational map indexed by structure, not sensory features).

Behrens, Muller, Whittington et al. (2018). “What is a cognitive map? Organizing knowledge for flexible behavior.” Neuron, 100, 490-509. High-profile synthesis. Proposes hippocampal-entorhinal system implements a general relational geometry engine. Abstract-map view enters mainstream neuroscience.

Whittington, Muller, Mark, Behrens et al. (2020). “The Tolman-Eichenbaum Machine: Unifying space and relational memory through generalization in the hippocampal formation.” Cell, 183, 1249-1263. Computational unification. Spatial cognition = special case of general relational learning. Named to honour Tolman (1948) and Eichenbaum & Cohen (1993). The abstract-map view becomes the leading theoretical framework (~2020 = mainstream).

Note: a 2026-03-22 Euronews article reported this work as a “breakthrough discovery.” This was prize journalism marking Doeller’s 2026 Leibniz Prize, not a new paper. The underlying science dates to 2010-2020.

REE did not derive its hippocampal architecture from this literature – the convergence is independent, from the architectural design direction. This constitutes convergent evidence that the hippocampal relational-trajectory-simulator framing is correct.

Source: docs/thoughts/2026-03-24_COHERENCE_MULTICONSTRAINT_HIPPOCAMPAL_NAVIGATION_CONVERGENCE.md

ARC-007
ARC-018
ARC-003
ARC-004
INV-006
MECH-022
MECH-033
Q-011

References / Source Fragments

See also: vmPFC — the complementary substrate that loads affective and normative content into hippocampal state-graph nodes at evaluation time. Hippocampus provides graph structure; vmPFC provides what the nodes contain (ARC-035).

docs/processed/legacy_tree/docs/architecture/hippocampal_braid.md
docs/processed/legacy_tree/architecture/Hippocampal_braid.md
docs/thoughts/2026-02-08_control_plane_modes_responsibility_flow.md
docs/thoughts/2026-02-09_starting_with_sensory_streams.md
docs/thoughts/2026-02-09_e2_hpc_interface.md
docs/thoughts/2026-02-12_DEPRESSIVE-PATH-PRUNING-HIPPOCAMPAL-ROLLBACK.md
docs/thoughts/2026-03-24_COHERENCE_MULTICONSTRAINT_HIPPOCAMPAL_NAVIGATION_CONVERGENCE.md