tinyHippo

A bio-plausible spiking model of the rat hippocampus in NEST (Izhikevich neurons), built to study memory consolidation through bidirectional replay and sharp-wave ripples across the full entorhinal–hippocampal loop:

EC LII → DG → CA3 → CA1 → EC LII/LV → mPFC
   ↑                              │
   └──────────────────────────────┘

The model scales from a 1% test network (~8k neurons, runs on a laptop) to the full rat hippocampus (~780k neurons, MareNostrum 5), with the same code path.

A companion document, hippocampal_timescales_as_circuit_spec.md, takes this model's validated timescales and mechanisms (synaptic tagging and capture, replay-driven consolidation) and specifies them as a memristor-based neuromorphic circuit — organic volatile devices for the hippocampal tag, inorganic non-volatile crossbars for the cortical store, with replay as the handoff between them.

What the model does

Capability	Flag	Status
Bidirectional replay — forward & reverse sequence replay during SWRs	(default)	validated at 12% (ρ_fwd +0.70, ρ_rev −0.54)
CA3 recurrent feedback loops — Watson 2025 four-way asymmetric wiring + E↔I	(always on)	core
DG pattern separation — granule sparse coding via basket feedback	--dg	validated at 12% (2.2% active/pattern)
CA3 pattern completion — auto-association, partial cue → full pattern	--pattern-completion	validated at 12% (sharp ~30% threshold, ablation-controlled)
Synaptic tagging & capture — early-LTP → tag → PRP → late-LTP on CA1→EC	--stc	implemented
Memory consolidation — hippocampo-cortical loop (EC LII, EC LV, mPFC)	--ec-lii --ec-lv --mpfc	implemented
Synaptic homeostasis — sleep downscaling, cortical L-LTP exempt	--homeostasis	implemented
Topographic DG wiring — clustered EC LII→GC fan-in, restores pattern identity washed out by uniform random sampling	--dg-ec-cluster-sigma	validated at 1% and 12% (small, real effect)
Pattern discrimination probes — per-population Jaccard identity/timing separation across interleaved patterns	--n-patterns	implemented

An interactive map of all capabilities, with the functions implementing each, is in capability_map.html (open it in a browser).

Quick start

Install NEST ≥ 3.9 (see INSTALL.md), then:

# 1% test network: replay + dentate gyrus, ~1 min on a laptop
python replay_scaled.py --scale 1 --dg --dg-scale 2 --no-figures

# CA3 pattern completion probe (intact vs ablated recurrence)
python replay_scaled.py --scale 1 --pattern-completion

# full consolidation stack
python replay_scaled.py --scale 1 --dg --ec-lii --stc --n-swr 3

Results are written to a self-describing HDF5 file (all populations, spike times, rates, and per-capability metrics), so analysis and plotting run anywhere without NEST.

Running on HPC

run.sh is the SLURM launcher (MareNostrum 5, MPI + OpenMP):

# replay + DG at 12%
sbatch --export=ALL,SCALE=12,DG=1,NO_STC=1,EC_LII=0,EC_LV=0,MPFC=0 run.sh

# pattern completion at 12%
sbatch --export=ALL,SCALE=12,PATTERN_COMPLETION=1 run.sh

# full stack with DG + consolidation
sbatch --export=ALL,SCALE=12,DG=1,N_SWR=14 run.sh

Single-neuron f-I calibration has its own job: sbatch --export=ALL,PROBE=dc run_calibrate.sh → run_calibrate.sh.

Layout

File	Purpose
replay_scaled.py	Main simulation — all populations, capabilities, and HDF5 export
tiny.py	Shared helpers (seeding, theta and SWR generators)
nest_dg_ca3_fi_calibration.py	DG/CA3 single-neuron f-I + DC-rheobase calibration
run.sh / run_calibrate.sh	SLURM launchers
plot_pattern_completion.py, replay_plot.py, make_paper_figure.py	Offline plotting from HDF5
reconstruct_connectivity.py / run_reconstruct.sh	Extract a projection's actual wired connectivity without a full simulation
capability_map.html	Visual capability reference
hippocampal_timescales_as_circuit_spec.md	Companion hardware doc — memristive circuit spec derived from this model's timescales

Results

Memory consolidation is dissociable from replay. Blocking late-LTP capture (--prp-threshold 999) leaves replay quality identical (Δρ_fwd = 0.000) while cortical consolidation goes to zero — separating the replay mechanism from the consolidation mechanism in the same model. Full circuit, 12% scale:

CA3 performs pattern completion. A partial cue of a stored assembly is restored to the full pattern by the recurrent collaterals, with a sharp attractor threshold near 30% cue. Ablating the within-group recurrence (sup_local = 0) abolishes it, confirming the collaterals — not the cue — do the work:

Sparsifying the cortex makes consolidation selective — but not yet specific. A cortical sparsity retune (§11 in RESULTS.md) turns saturated, all-cell L-LTP into a differentiated trace (7.2% of EC cells consolidate, weight CV 0.17). Scored against standard engram criteria (Josselyn & Tonegawa 2020), that trace is sparse and persistent but not yet specific: no cortical population discriminates pattern A from pattern B (§13), and hippocampal lesion does not impair cortical recall at the 12% scale tested (§12, Test 3 negative). Two convergent, non-topographic projections are implicated — DG's perforant path and, more severely, CA3→CA1 Schaffer collaterals at 100% density — and a topographic fix (--dg-ec-cluster-sigma) moves DG's identity signal off zero for the first time, though the effect is still small (§17–18). Full scorecard and in-progress work: RESULTS.md.

Key references

• Watson et al. (2025) Cell Reports 44:116080 — cell-specific CA3 wiring (superficial/deep split)
• Marr (1971); Nakazawa et al. (2002) — CA3 auto-association / pattern completion
• Frey & Morris (1997) — synaptic tagging and capture
• Kassab & Alexandre (2018) — DG mossy-cell threshold classes
• Andersen et al. (2007) The Hippocampus Book — reference neuron counts
• Josselyn & Tonegawa (2020) Science 367:eaaw4325 — engram criteria (sparse, persistent, specific, sufficient, necessary)
• Dolorfo & Amaral (1998) — entorhinal-dentate medial-lateral topography, motivating the clustered perforant-path fan-in
• Izhikevich (2006) — polychronization / delay-based temporal coding
• Caus, Sławek, Mazur, Zawal, Baś, Szaciłowski, Talanov & Abdi (2026) — memristive hippocampus hypothesis; candidate tag-element devices for hippocampal_timescales_as_circuit_spec.md