NDM (Neuroplasticity-based Decision Memory) · the substrate

Memory that
learns the work.

NDM (Neuroplasticity-based Decision Memory) plugs into the systems your business already runs on — email, Slack, ticketing, shared drives — and self-learns the business logic your senior people apply every day. Every adopted rule is traceable to its source, auditable, and reversible. The same memory is queried by your team directly and navigated by the AI agents acting on their behalf. Instant value across finance, legal, medical, claims, compliance — anywhere business rules sit over a structured ontology. Pair with the neuroplast ontology builder to start from a turnkey domain model on day one.

Domain-ready · finance · legal · medical · compliance Plugs into · email · slack · drives · ticketing Two consumers · your team + the AI agents acting for them
What NDM is

Not a model. Not a vector store. A substrate.

The hard problem in regulated decision support is not generating text. It's representing accumulated judgment in a way you can interrogate, audit, and update without retraining. NDM solves that, three layers deep.

01 — Representation

A bitemporal knowledge graph.

Every claim is a node. Every relationship a typed edge. Every fact has a valid_at (when it was true in the world) and a tx_at (when the system learned it). Memory has a shape, not just a token count.

02 — Lifecycle

Proposed → adopted → superseded.

New facts enter as proposed. Repeated corroboration promotes them to adopted. Contradictions don't overwrite — they supersede, leaving the audit trail intact.

03 — Rewiring

Sleep tick. Hebbian. Pruning.

A periodic consolidation job replays activity, strengthens co-fired patterns (Hebbian), runs PageRank + Louvain to surface centrality, decays stale rules, and prunes evidence-poor edges. The substrate gets sharper over time, not bigger.

Memory lifecycle

One claim. Five states. Always recoverable.

Click any stage — or press play — to walk through how a single claim moves through NDM. The audit log preserves every transition. Pruned never means gone.

Stage 01 · Proposed

First observed.

A passive extractor watches a memo or a Slack thread, proposes a claim, and tags it with the source span. Nothing is acted on yet. NDM can see it; it just can't cite it.

Proposed CM-04102 · v1

"Cap exposure 20–25 % below request when top-3 concentration exceeds 60 %."

Confidence0.31
Sources1 memo · M. Aravena
Valid_at2026-02-14
Cite-eligibleNo — below adoption threshold
Auto-advances every 2.4 s · click a stage to jump
Eight mechanisms · from neuroscience

The substrate borrows from how brains actually learn.

Each mechanism is implemented as a discrete primitive — testable, citable, parameterized against calibration sets. None of it is metaphor.

01

Hebbian co-firing

Patterns that fire together strengthen their connection. The substrate learns "this signal pair predicts this decision" without supervision.

02

Time-based decay

Rules with no recent corroboration fade. Confidence is a function of evidence freshness. The substrate forgets the same way experts do — slowly, in the background.

03

Supersede > delete

When a new claim contradicts an old one, the old one is marked superseded — never erased. The audit trail survives the update. You can always rewind.

04

Sleep-tick consolidation

A periodic job runs PageRank + Louvain across the memory graph, clusters co-fired patterns, and surfaces emergent rules. Sharper, not bigger.

05

Critical-period gating

New NDM deployments start in a high-plasticity phase — rules form fast. Once enough are corroborated, plasticity dampens. Mature memories don't get overwritten by noise.

06

Evidence pruning

Edges with thin or stale evidence are dropped, freeing the substrate from low-signal connections. The graph stays interpretable as it grows.

07

Hybrid retrieval

BM25 + dense vectors + graph centrality. Each query lights a path through the substrate; NDM cites every node it walked.

08

Bitemporality

Every claim has two timestamps: when it was true in the world (valid_at) and when the system learned it (tx_at). Audit questions like "what did NDM know on date X" become a single query.

The substrate · stack view

Six layers. One audit log.

Every action that touches memory goes through this stack. Each layer is pluggable; each layer is observable. The audit log threads through all of them — every read, every write, every retraction.

Layer 1

Sources

Memos, decisions, chats, forms, policies, outcomes. NDM ingests what your senior people already produce — no behaviour change required.
PDF · DOCX · MD
Slack export · Webhooks
Layer 2

Ingestion

Parsers strip layout, preserve span offsets, and hand documents to an entity / relation extractor that proposes claims with source citations.
Docling · PyMuPDF
LlamaIndex · CASDK
Layer 3

Memory

The bitemporal graph. Every node is a claim or a pattern; every edge a typed relationship. valid_at and tx_at on every fact. Supersede > delete.
Neo4j Enterprise · GDS
SQLite event ledger
Layer 4

Retrieval

BM25 + dense vectors + graph centrality. Hybrid scoring picks the nodes most likely to inform the decision; cites every one.
neo4j-graphrag
text-embedding-3-small
Layer 5

Consolidation

The sleep tick. Replays activity, runs PageRank + Louvain, fires Hebbian updates on co-activated patterns, decays the stale, prunes the thin.
GDS · scheduled cron
calibrated thresholds
Layer 6

Reasoning layer

A LangGraph workflow walks the substrate, drafts a recommendation, names confidence, and surfaces every cited node. Consumed by your team directly (chat, dashboards) and by downstream AI agents (via MCP, API). Your team decides — the reasoning layer argues.
LangGraph 1.x
Claude Agent SDK · MCP
Spans all layers

Audit ledger

Append-only, hash-chained. Every memory operation — propose, corroborate, adopt, supersede, prune — is a recorded event with provenance. Regulator-ready by construction.
SQLite WAL · JSONL
schema-versioned
Where NDM sits

Not RAG. Not fine-tuning. Not context engineering.

Each existing pattern handles part of the problem. None handle accumulated judgment in regulated environments. That gap is what NDM closes.

RAG / vector store Fine-tuning Context engineering NDM
Adds new knowledge without retraining Yes No Yes Yes
Distinguishes proposed vs adopted No No No Yes — first-class
Bitemporal — answers "what did it know on date X" No No No Yes
Audit log of every memory change Partial No No Yes — by construction
Reversibility — supersede > delete No No No Yes
Captures judgment, not just facts No Implicit No Yes — patterns first-class
Cost to add a corrected rule Re-embed, re-index Re-train Edit prompt One supersede transaction
Talk to neuroplast

Bring your domain. We bring the substrate.

A thirty-minute scoping call. We sketch what NDM looks like running on one of your regulated decision flows — credit, claims, radiology, compliance, vendor risk, or your call.