braineon · the research brief · v1.0 · braineon.ai/research
braineon is an engineered memory system. It follows the encode → consolidate → retrieve → monitor arc your field has characterised for a century — so here it is, mapped term by term, every claim cited, and the places the analogy breaks stated plainly.
What this is, and is not. This is not a brain model and makes no mechanistic claim. Where we write “like the hippocampus,” we mean functionally analogous — a shared vocabulary, not an equivalence. Several of the most familiar labels — consolidation, reconsolidation, encoding — are places the engineered system diverges from biology rather than reproduces it. We say so at each one. The divergences are the point.
Scope: functional analogy · Numbers: none (benchmark section is methodology only) · Sources: 28 cited below · Status: draft, open to correction at researchers
Each construct as memory science defines it, its counterpart in braineon, and — honestly — whether that counterpart improves on, diverges from, or is only loosely analogous to the biology. Superscripts link to the References.
In memory science, the human counterpart of abstention is a control decision gated by monitoring: a feeling-of-knowing decides whether to even search, and a recognised retrieval failure terminates the attempt.8 braineon implements the same shape — measured signal, explicit threshold, suppressed low-confidence output — as an engineering discipline rather than a felt one.
The formal frame is selective prediction — a predictor paired with a gate that answers only when a confidence signal clears a threshold τ. Lowering τ raises coverage (how much is answered) and raises selective risk (error among answered); raising it does the reverse. The oldest version is Chow’s reject rule; the modern treatment is the risk–coverage curve.1819 The floor is not hand-picked: it is set on held-out data to hold a target selective risk.
The confidence signal is a composite, and each part is a real technique:
The honest boundary. “Zero confabulation” is a guarantee about the store, not automatically about the output. The deterministic store cannot invent a memory; the meaning-aware retrieval layer, like any such system, can paraphrase or blend. The abstention model is precisely the guardrail — source-binding plus a faithfulness gate — that carries the store’s integrity through to the answer. That is the engineered counterpart of intact source monitoring,10 not a claim that generation is magically incapable of error.
Biology conflates two things braineon keeps strictly apart: a memory fading and a memory being gone. The decay model is two mechanisms, not one.
Decay of priority (soft, reversible). Disuse lowers a memory’s retrieval priority; use raises it — the continuous form of the LRU/LFU intuition from cache design. A salience weight combining recency and frequency multiplies the retrieval score, so stale items sink without being lost. The retention shape is configurable: it can be fit to an Ebbinghaus-style curve, or to a per-item half-life learned from use — the approach Duolingo’s half-life regression made concrete for spaced repetition.21 The curve is a setting, not a fate.
Erasure (hard, irreversible, audited). Distinct from decay: a deleted memory is not demoted, it is unrecoverable. The mechanisms are ordinary and provable — tombstones, retention policies, and crypto-shredding (encrypt each item under a per-subject key; destroy the key and the ciphertext is permanently unreadable), the standard way to satisfy the GDPR Article 17 right to erasure in append-only stores.22 This is the affordance biology cannot offer: brains cannot reliably delete; braineon can, and can prove it (see the erasure audit in §4).
Deterministic consolidation, in this frame, is just the rebuild of derived state (index, embeddings, weights) from the source files — a materialized view, reproducible byte-for-byte from the same inputs.23 Where biological consolidation is lossy and unrepeatable, this rebuild is idempotent and re-runnable. It is the same word for a deliberately different thing.
Methodology only. This section describes how braineon should be measured. It reports no results. Until a pre-registered protocol is run on held-out data, no benchmark number exists and none is claimed here. When numbers come, they will arrive as full curves with confidence intervals — never a single hero figure.
Retrieval quality, the way the field scores memory. Recognition decisions are scored by signal-detection theory — sensitivity d′, criterion, and a full ROC/AUC swept across thresholds24 — and the retrieval side by precision@k, recall@k, MRR and nDCG.25 We report the whole curve on a fixed, gold-labelled query set, not one operating point.
Calibration — does stated confidence mean what it says? A reliability diagram plus Expected Calibration Error, Brier score, and negative log-likelihood over a held-out stream of (stated confidence, was-correct) pairs.20 This is the evidence the metamemory claim in §1 owes you; without it, “calibrated” stays a hypothesis.
Abstention — is silence earned? The risk–coverage curve and the area under it (AURC): rank queries by confidence, walk the threshold, and plot error-among-answered against fraction-answered.19 This is the metric that credits a well-placed “I don’t know” instead of punishing it.
Decay and erasure. Fit the retention curve at the item level (exponential vs power-law, reported with goodness-of-fit) rather than over-averaging.13 Then a separate erasure audit: plant canary memories, delete them, and run an adversarial recovery battery (exact, paraphrase, partial-cue, embedding-neighbour) to confirm zero recovery. A merely-decayed item must fail this audit. (This audit is a braineon-proposed protocol, not an established field standard.)
Provenance. Every answer’s citation is checked for support: human-judged AIS (“Attributable to Identified Sources”) on a stratified sample,26 citation precision / recall / F1 at scale,27 with an automated entailment check calibrated against the human labels.
Guardrails. Pre-register the protocol and primary endpoint; hold out test data with contamination checks; anchor against human baselines where apt; report bootstrap confidence intervals and n on every metric.
The affordances a brain can’t offer — and the honest open questions where biology still wins.
Primary literature for every cited claim above. Neuroscience and cognitive-science sources are canonical; engineering sources name the technique each model corresponds to.
How this brief was made. It was drafted through an adversarial, multi-perspective process — a neuroscientist, a skeptic, a systems engineer, a metamemory specialist and an evaluation methodologist each pressed the same claims from a different angle — and then checked by an independent review that did not write it. Every neuroscience claim above binds to a cited source; contested points are flagged as contested. If you study human memory and something here is wrong, imprecise, or overstated, we want to hear it.
This brief is a draft held open to scrutiny. Bring your corrections, your citations, and the divergences we missed.
Scrutinise the mapping →