Recall finds what would come to mind, and Continuum says so when nothing does.

A model answers anything it's asked, from whatever its training left behind. Continuum puts the recalled facts in front of it with one rule: answer from these alone, and say so when they don't answer. The review pipeline's prompt puts it exactly: "If they don't answer the question, reply exactly: Memory doesn't hold the answer."

On 28 September the first experiments asked, among their questions, who won the 1998 FIFA World Cup, which nothing in the memory knows. The closed-book baseline, the same model with no memory, answered France. Every method backed by the memory, by words, by meaning, fused or by activation, came back with the memory not holding the answer. One question is a single reading, and the probe sets grow from here.

Activation adds a line of its own. Under it, a fact is retrieved only above the retrieval threshold, τ, and a fact below it drops into belowThreshold, returned beside the answer, so "why wasn't this recalled?" always has a reply.

Activation is ACT-R's equation for how easy a memory is to reach right now, computed for every candidate at the moment of the question and never written back: A = min(B + Σ W·S + ε, τ + band). The activation series tells the science; here is what each term does in Continuum.

• Base level, B. ln(β + Σ t^−d) over the ages of a fact's references, with d = 0.5, the decay Anderson and Schooler measured and flashcard schedulers found again. References older than the window are counted by Petrov's tail, which is ACT-R's own optimised learning. The floor, β = 0.002, was fitted in the evaluation harness, and it is what keeps a faded fact reachable by a strong enough cue. Until answers cite facts, a fact's one reference is its creation.
• Spreading, Σ W·S. What is in play lends strength to what it connects to. In fact recall the question is the one cue, and lends by meaning, scaled up to S_max = 4. In the design, the graph lends too: a declared connection lends S_max − ln(fan + 1), so a vague cue with many connections lends less than a precise one, and it halves with every hop, which is Soar's multi-hop design where ACT-R spreads a single step.
• The ceiling, τ + band. However often a fact is recalled, it can rise no higher than its evidence: the kind of author who asserted it (a person 4, an agent 2, the system 1.5), plus 1 if it is grounded, 1 for each further independent author up to 2, and up to 2 either way for its record of being right. It is a cap, never an added weight, so repetition can't buy authority.
• The threshold, τ = −4.0. A fact is retrieved above it and not below, unless the validity gate has already excluded it. τ is the value the experiments are calibrating.

Every fact comes back with its score taken apart: similarity, base level, spreading, the raw sum, the ceiling, the final score, whether the ceiling applied, whether it is contested, and, under hybrid, its rank in each list and its fused score. The playground shows the lot, and lists the facts below the threshold under "why these weren't retrieved".

A chat turn recalls up to eight facts and hands them to the model in one system message: "Facts recalled from this person's memory in Continuum for their message, strongest first. Use the ones that answer it, say so when none do, and state nothing as fact that isn't here or in the conversation." Then the statements, numbered. What comes back is grounded, and it reads like a pick from a list: the one statement that answers, repeated, where an answer should put several facts together into what they mean.

The model has four jobs with recalled facts: pick the ones that answer, combine several into one reply, qualify what they leave open, and refuse when none answer. Refusing works, and combining is the problem the experiments are for. Tristan, on 27 September: "If we are supplying questions, facts, and the user's question... What exactly is the LLM's purpose?"

The experiments come at it from three sides. From the memory's side, a fact is an index and its section is the payload, so the grounded harness, continuum model-quality grounded, can hand the model the same facts five ways: none, plain, structured, scored, or with their source sections attached. From the graph's side, the design resolves facts that say the same thing into one, so a question reaches the facts around its answer as well as the answer. From the answer's side, each variant in the review pipeline can pin its own chat configuration, so a prompt or a model is a variant as a retriever is.

An experiment is questions, one per line, with ! marking the ones the memory can't answer, crossed with variants: a method each, and one with none, the closed-book baseline. Up to 50 questions and 12 variants make an experiment, each pairing a trial, and the experiment runs as a Temporal workflow on the evaluation queue.

A trial retrieves with its method, assembles the prompt from what came back, each entry cut at 2,000 characters, and asks the production chat model for the answer. Then it checks the answer: that it finished and isn't only reasoning, that every citation like [2] points at an entry it was given, and whether it abstained. Each step is traced, and the Dashboard lays a run out as a tree on a timeline. A person rates each answer good or bad beside the scoreboard.

Base level decays over a fact's history of use, and that history is a record the other stores can't rebuild. The event plane is ClickHouse's context_events: an append-only receipt per thing that happened to a fact, Used, Corroborated, Rejected, Vindicated, Refuted or Traversed, with no content column, kept 180 days. In the design, a Used receipt is written when an answer cites a fact, and activation reads the windows of receipts as each fact's references.

Two kinds are left out on purpose. Created is read from the fact itself. Presented must never be recorded, because being shown a fact is not evidence of using it: the first reinforcement loop recorded every fact recall returned, strengthened whatever the retriever already favoured, and was removed a day after it was built.

The first design was a cascade: Qdrant first, then graph expansion, then scoring over what they found. Recall was bounded by whatever Qdrant returned first, and a fact that meaning missed could never be scored at all, which is exactly how an assistant forgets the rule it was given. So candidates come from every source at once, pooled, fused by rank and scored once. The pool is two lists, not three: with statements and questions as two dense lists beside one of words, meaning outvoted the words and lost "never".

A learned ranker needs labelled relevance data, and there was none. ACT-R gives a prior fitted against human memory over four decades, with every term named, which is what makes "why wasn't this recalled?" answerable. It still has to win: what activation adds to the fused ranking is measured on the same questions, and it stays for what it adds.

Activation depends on now: the decay term is a function of the time since each reference. A score written back to the row is stale the moment after the write, and it would alter a fact from a model that may later be found wrong, which the context store's first law forbids. Computed at query time, a corrected constant rescores the whole memory on the next question, with no migration.