Developmental pruning and sparse memory cognifold

Date: 2026-06-09 Status: Compass (V4/V5; off the V3 critical path) Claims: MECH-362, Q-057 Origin: thought-intake thought_intake_2026-06-06_ca3_development_sparse_structured_connectivity.md

This is a developmental architecture compass, not a V3 implementation target. The two claims registered here are substrate_conditional / implementation_phase: v4 – they are promote/demote-suppressed and do not imply an experiment is owed on the current substrate. Do not build a developmental pruning substrate or queue an experiment from this doc without an explicit decision routing it onto a version’s critical path.

1. Empirical anchor (verified)

Vargas-Barroso V., Watson J. F., Navas-Olive A., Schlogl A., Jonas P. “Developmental emergence of sparse and structured synaptic connectivity in the hippocampal CA3 memory circuit.” Nature Communications (2026), article s41467-026-71914-x (ISTA). https://www.nature.com/articles/s41467-026-71914-x

  • Multicellular patch-clamp circuit mapping of up to 8 CA3 pyramidal neurons at three postnatal stages: P7-8 (early), P18-25 (adolescent), P45-50 (adult).
  • CA3 recurrent connectivity transforms local / dense / strong / near-random -> distributed / sparse / structured.
  • Early: a single synaptic event can be sufficient to trigger a postsynaptic spike.
  • Mature: spatial summation of several (weaker) inputs is required to drive a spike.
  • Authors’ framing: “tabula plena” – a full, over-connected starting substrate refined by pruning – rather than tabula rasa; the sparse/structured endpoint may emerge via experience-dependent mechanisms.

Cross-species corroboration that the mature endpoint is structured (not random): Tang et al., “Human hippocampal CA3 uses specific functional connectivity rules for efficient associative memory,” Cell (2024).

2. The two registered claims

MECH-362 – Subtractive developmental sparsification

Mature sparse/structured memory connectivity emerges by pruning / down-weighting an initially over-connected, dense, near-random substrate (“tabula plena”), not by additively growing connections onto a sparse skeleton. The corollary at the recall stage: mature recall requires convergent summation of several weak inputs rather than a single strong cue.

Distinct from MECH-120 (SWS synaptic homeostasis / Tononi SHY). MECH-120 is nightly homeostatic down-scaling of synaptic weight to restore signal-to-noise; MECH-362 is a one-way developmental trajectory of recurrent connectivity from dense-random to sparse-structured. They may share machinery (down-weighting), but the timescale, directionality, and function differ.

Relation to ARC-019 (staged developmental curriculum). REE’s developmental claims (ARC-019, IMPL-019) are currently additive – capability is added stage by stage. MECH-362 proposes a subtractive stage: an over-connected exploratory phase that is pruned. If adopted, MECH-362 most naturally amends ARC-019 rather than standing fully alone.

Pruning is not (only) deletion. Down-weighting, gating, or residue-tagged de-authorization are plausible computational analogues – see Q-057.

Q-057 – Exploratory-overconnected vs mature-sparse substrate

Should REE distinguish an early over-connected exploratory developmental substrate from a mature sparse/structured substrate – and is developmental sparsification best modelled as deletion, down-weighting, gating, or residue-tagged de-authorization? Sub-questions:

  • Does REE-v3’s cue-authority problem partly reflect immature single-cue authority rather than mature convergent-weak-input recall?
  • Can offline integration act as pruning / contextualisation rather than merely consolidation?
  • What developmental gates should exist before a memory trace can influence action release?

3. Three-phase developmental schedule (hypothesis)

  1. Over-connected exploratory phase – many weakly-constrained associations, high plasticity, broad coupling, low action authority.
  2. Pruning / selection phase – repeated offline + waking experience down-weights unstable / non-useful pathways.
  3. Sparse structured retrieval phase – mature recall requires coordinated multi-cue activation rather than single-cue domination.

Recall formulation (from the intake):

mature_recall_activation = f(convergent_weak_inputs, context_match, residue_state,
                             goal_state, self_world_tag, precision_gate)

rather than max(single_strong_cue).

4. Where the V3-relevant strand already lives

The convergent-weak-input vs single-strong-cue corollary is the one strand with present-day V3 relevance (REE-v3 currently exhibits single-representative over-authority). It is folded in as a diagnostic lens – NOT a separate V3 claim – in the selection-authority plan-of-record: modulatory_bias_selection_authority_design.md “Developmental framing (compass)”. Registering it as its own V3 claim would duplicate the active selection-authority / rule-apprehension claims (MECH-314/341, ARC-062).

  • ARC-006 – entities are sparse, persistent, bindable structures (the sparse endpoint).
  • ARC-007 – hippocampal map substrate (biological locus is the CA3 recurrent circuit).
  • ARC-019 / IMPL-019 – staged developmental curriculum (currently additive; MECH-362 would add a subtractive stage).
  • MECH-120 – SWS synaptic homeostasis (SHY); closest existing down-scaling analog (V4).
  • MECH-094 – provenance / hypothesis tagging (failure mode: confabulatory completion from un-pruned over-connection).
  • MECH-076 – attractor lock-in (failure mode: belief fixation / un-pruned early attractor).
  • Play-mode cluster (ARC-049/050, MECH-194-199, INV-058/059/060) – the natural home for an “exploratory over-connected phase,” but substrate-blocked in V3 (no play_frame_tag); do not queue against it.

6. Cautions (carried from the intake)

  • Do not overgeneralise mouse CA3 development to human cognition or REE design.
  • Do not treat dense early connectivity as automatically good/bad.
  • Do not assume pruning = deletion only.
  • Do not make this a V3 implementation target without a specific substrate gap.
  • Guardrail: if a future agent tries to convert this into “add more connections” or “delete weak connections,” stop and reframe. The correct extraction is: mature sparse structured memory may emerge through developmental pruning of an initially over-connected substrate.

REE is developed by Daniel Golden (Latent Fields). Apache 2.0.