AI memory that refuses to guess.

01 / minimum sufficient retrieval

Retrieve less. Know more.

Most AI memory systems either forget too much or retrieve too much. Neuroflow uses progressive retrieval: it starts with the smallest relevant memory object and descends toward evidence or raw source only when the task requires it.

02 / failure analysis

Same query. Different failure mode.

why it matters · Neuroflow preserves modality, evidence and source traceability, so a hedged statement can never harden into a commitment on the way into memory.

03 / memory architecture

The Memory Map — six layers, one chain of custody

Summaries help the system navigate. They are not evidence and cannot prove a claim. When a claim matters, Neuroflow descends to evidence and raw source.

04 / retrieval controller

The Retrieval Controller decides how much memory is enough

For every task the controller chooses an expansion depth — from no memory at all to raw source — then applies policy, sensitivity rules, expansion limits and audit requirements before returning a context bundle.

05 / memory lifecycle

From source to durable memory

1. 01

   ingest

   Documents hashed, versioned, stored immutably. No interpretation at this stage.

2. 02

   extract

   Verbatim spans lifted with source offsets and a faithfulness score.

3. 03

   atomise

   Each span becomes at most one neuron carrying exactly one claim.

4. 04

   assemble

   Neurons grouped into nuclei shaped by the questions the business asks.

5. 05

   audit

   Auditor agents admit, split, or reject before anything becomes retrievable.

6. 06

   revise

   New evidence supersedes old neurons; history is retained, never overwritten.

fig. 04 · layer assembly

06 / memory firewall

Memory firewall: what gets blocked

Neuroflow does not rely on model opinion to decide whether memory is safe. Guardrails are deterministic checks with fixed verdicts and logged reasons.

idconditionverdict

G-01unbacked claimno evidence span resolvesreject

G-02modality drifthedged source rendered as firmreject

G-03multi-claim neuronatomicity violatedsplit

G-04stale evidencesource superseded by newer docquarantine

G-05summary cited as factorientation layer is non-citablereject

G-06cross-account leaknucleus outside cluster scopeblock

fig. 03 · rejection gate

07 / auditor agents

Auditor agents run the factory floor

Nothing enters memory unaudited. Each layer has a dedicated auditor with the authority to reject.

$ audit --layer=evidence --batch=0f21
span#0142 offset match ......... ok
paraphrase drift 0.00 ......... ok
modality preserved ............ ok
→ 412/412 spans faithful

$ audit --layer=neurons --strict
claims without evidence ....... 0
multi-claim neurons ........... 0
unsafe generalisation ......... 0
→ 212/212 neurons admitted

$ audit --layer=nuclei --coherence
cross-topic contamination ..... 0
coherence score ............... 0.94
orphan neurons reassigned ..... 3
→ 14/14 nuclei coherent

08 / case study

US Dairy Commodities — Milky Way Co. vs. ACME

Three emails. Hover a message to follow the chain of custody: evidence extracted, neuron created, nucleus assigned.

09 / pluggable access

How memory is reached — and what an AI app receives

Neuroflow returns a context bundle, not a pile of chunks: the minimum sufficient text, the trace that proves it, the expansion depth used, and a verdict.

fig. 05 · inspection surface

apiGET /neurons?nucleus=…

Every response ships with evidence spans and source ids.

mcpneuroflow.retrieve

Agent-native access with the same guardrails applied.

reviewhuman admit / reject

Auditor verdicts are reviewable and reversible by an operator.

what an AI app receives

context bundle

The customer discussed approximately 3,000 MT of Q3 butter, but the source explicitly says the volume is indicative and not a firm commitment.

trace

email#3 → span#0142 → evidence → neuron → nucleus

expansion level

L2 nucleus + L4 evidence

verdict

answerable with evidence

10 / operating profiles

Memory quality starts with intent

Neuroflow brains are initialised with an operating profile. A legal memory brain, a writing memory brain, a research brain and an industrial telemetry brain should not form the same neurons. The profile declares what matters, what counts as evidence, what must be retrieved fast, and what must never be inferred.

The plugin is general. The brain is specialised.

11 / positioning

Not another vector database

Vectors help Neuroflow find candidates. They do not decide truth. Truth lives in the evidence chain: raw source → source span → evidence item → neuron → retrieval trace.

12 / operating principles

Four rules the memory layer never breaks

fig. 01 · ingest line

P1

Source authority

Every memory must resolve to an immutable source span. If the chain breaks, the claim is dropped, not softened.

P2

Modality preservation

“May”, “indicative” and “subject to” are not writing style. They are business meaning, and they are stored verbatim.

P3

Atomic memory

One claim per neuron makes memory contestable, replaceable and auditable without rewriting its neighbours.

P4

Safe refusal

When evidence is missing, Neuroflow returns the gap instead of inventing an answer. A missing memory costs a follow-up question; a false memory costs a contract.

13 / questions

What people ask first

Is this a vector database?

No. Vectors are one index among several. Truth lives in the evidence chain, not in embedding proximity.

What happens when sources conflict?

Both neurons stay. The conflict is surfaced as a contested claim rather than resolved silently.

Can it answer without evidence?

It will not. The system returns the gap and the closest grounded context instead of a guess.

How is memory kept current?

New documents supersede prior neurons; superseded memory is retained with its retirement reason.

14 / definition

Neuroflow is a pluggable memory manager for AI systems.

It preserves raw sources, extracts atomic evidence, builds durable memory objects, applies guardrails, and retrieves the minimum sufficient context for each task. It is designed for systems where a missing memory is acceptable, but a false memory is toxic.