Reading memory

Ordinary Markdown excerpts carry a derived prose qualification. Hypothetical headings, quotes and speaker turns cannot become unqualified factual evidence merely because an excerpt omits its heading. English/Russian view inference and independent answer-language requests are described in knowledge language and discourse.

Akno separates evidence discovery, direct answering, exact reads, relationship inspection, time-based lookup, and automatic host context. The split keeps fast or inspectable operations from silently becoming generative.

Do not call both answer and recall by default for one question: answer already retrieves. Call recall separately when a person or agent also needs to inspect the evidence cards.

akno recall "When does the Zephyr QX-100 warranty expire?"
akno recall "invoice" --type receipt
akno recall "policy wording" --include source --depth full
akno recall "warranty" --source document --ownership orphan

Recall runs a staged pipeline:

The search arms use incompatible score scales, so Akno fuses ranks instead of comparing raw BM25, cosine, graph, or reranker values.

Recall chooses or accepts one mode:

Search mode and memory view answer different questions. Mode controls how Akno searches; memory_view controls which meaning of an Akno-managed retained item is eligible for the primary result:

Akno infers only high-precision cues such as “reported,” “planned,” “decision history,” “open question,” or “hypothetical.” An ambiguous lookup stays factual; an ambiguous explore request uses all. Pass --memory-view or memory_view when the host already knows the intent. The explicit value always wins.

An explicit question about an estimated date, delivery window, or when an item is expected to arrive selects planning, including equivalent Russian requests. Active tentative schedules and deadlines remain retrievable even when their retained kind is a claim or event. Their dates, reporter and uncertainty remain visible; they are not promoted to confirmed occurrence or canonical facts. Estimation of weight or cost alone does not change the view, and hypothetical or counterfactual claims do not become schedules.

Eligibility is applied to isolated managed-memory chunks before lexical, vector, graph, and reranker candidate limits. Ordinary authored text and document evidence remain eligible in every view. If a relevant retained item exists only outside the selected view, line-level recall returns it as contextual qualified memory with a note instead of claiming an empty absence; factual answering and automatic injection still refuse to use it as a fact. A missing, stale, malformed, or duplicate semantic projection reports partial_memory_index.

A result card reports role, relevance, contributing candidate arms, cited excerpts, and approximate budget. Question mode also reports concept coverage. A relevant page that does not cover one requested attribute is not evidence for that part of the answer.

While any canonical Markdown source is quarantined, recall carries source_conflict as typed degradation even when other evidence is still usable. An empty result therefore cannot be mistaken for proof that the conflicted source contained nothing relevant.

When a returned line is an Akno-managed level-one memory, it also carries memory qualification: kind, attribution, commitment, disposition, polarity, epistemic basis, and answer_eligible. Reports, hypotheses, counterfactuals, proposals, rejected decisions, plans, and questions remain searchable and readable with that status. They are not silently converted into ordinary current facts. Authored prose has no memory field; managed canonical claims and accepted decisions report answer_eligible: true. Time-scoped items also report their stored temporal envelope, a read-time clock_relation, whether they are actionable, and current_eligible. A past validity interval may remain valid historical evidence while being ineligible for a question about what is current. An owned marker whose current semantics cannot be parsed reports status: "unavailable" and answer_eligible: false; malformed or unmigrated managed memory never falls back to authored-fact behavior.

An eligible level-two payload carries a separate observation qualification with its stable id, exact subject, disposition, proof count, and every current leaf fact locator. It is labeled as derived in recall. Ordinary factual queries rank an observation-backed card below direct evidence; pattern, habit, recurrence, and tendency queries may promote it. Malformed markers and blocks whose facts, hashes, proof groups, subject, or placement authority no longer qualify remain readable with status: "ineligible", but are excluded from factual recall, automatic context, and graph traversal.

New observations that passed the separate evidence-scope gate also expose scope_assessment: {status: "assessed", fingerprint}. The fingerprint binds the exact sentence, complete bounded evidence context, prompt, and configured model; it is not a probability or certificate of world truth. Legacy observations omit this field and are never relabeled as having passed a check that did not exist when they were created.

Reflected L3 conclusions carry prose qualification independently of indexed fact rows: reflection_supported is factual and answer eligible; reflection_support_stale or reflection_lineage_missing is history and not eligible for a current factual answer. Explicit reads and pins keep that qualified history inspectable. Recall can return it as accurately qualified contextual evidence when no current evidence is available; automatic context and answers exclude unsupported reflection prose, including in the all view. A history recall query can inspect the old conclusion without claiming it remains supported. Other supported conclusions on the same page retain their eligibility. These bindings describe derived support, not authored claims or world truth.

A successful reranker may remove judged-irrelevant candidates. Candidates outside its bounded window are unjudged and are not used to fill holes the model never approved. If every judged candidate is rejected, recall returns honest empty.

If reranking fails or returns an invalid structured response, Akno preserves fusion order and reports typed degradation. It never applies an irrelevance filter from an unvalidated result.

akno answer "How long is the Zephyr QX-100 warranty?"
akno answer "How long is the Zephyr QX-100 warranty?" --context
akno answer "How long is the Zephyr QX-100 warranty?" --rerank

answer uses question-oriented recall, then asks the answer role for structured cited blocks. Akno validates the opaque evidence labels, checks introduced numbers and negation against cited text, and renders persistent locators itself. A separate verifier call judges each answer block only against its nested evidence.

For a bound original frame or a typed temporal record, a separate source-reading call identifies the selected predicates and their dates without seeing the new draft or question. These grounded readings constrain the verifier; it cannot rewrite their dates or status to fit the draft. They do not authorize additional facts from the original frame. Complete-record rendering compares every selected predicate, while focused answers may omit unrelated neighboring predicates.

For eligible observations and reflected conclusions, Akno expands the exact current support of each selected population. A separate source reading fixes the case counts and any explicit shared or disjoint identities without seeing the answer. Another reading fixes the answer’s own case counts and scope without seeing source records. A small comparison checks both fixed readings before the ordinary semantic verifier runs. Both must pass. Sharing dates does not combine two journeys and two workshops into two shared sessions; multiple facts, copied sources or competing dates for one identified event do not add cases. Exact repeated lineage is counted once, including when a draft cites both an observation and its raw support.

These checks use the configured answer role and its existing ceiling, without a new search or identity store. A source reading or comparison that cannot be grounded returns verification_unavailable; an established scope mismatch rejects the affected block, preserving other verified blocks. Finite examples cannot establish a broader habit, preference, motive or causal guarantee. These are source-support checks, not independent verification of reality.

An unavailable audit holds only its own block. Independently verified blocks may still be returned with partial coverage and answer_verification_failed degradation. Optional validation.unavailable_blocks counts blocks without trustworthy verification, separately from semantic rejection; this includes remaining unattempted blocks after a provider failure. If none survive, the result remains verification_unavailable with no answer or citations. The verifier does not retry or repair held blocks, reuse another block’s evidence, or label partial text complete. Private verdicts use concise wording under the caller’s unchanged model budget; every required predicate, role, scope and exact-quotation check still applies.

The verifier also compares each selected predicate with its own source date and status. A shared subject or sentence cannot transfer one action’s date onto another. For example, a dated booking and an undated completed payment can answer the booking date and payment completion; they cannot answer when the payment occurred. An answer can identify that missing date without denying the completed payment. Unavailable or ungrounded audits withhold the block; a negative or incomplete audit rejects it, even if the broader semantic verdict is positive.

Unsupported blocks are withheld. A missing model produces degraded/not_answered; complete empty recall produces empty/not_found. If equally applicable evidence gives incompatible values without an authority rule, Akno abstains rather than choosing one or inventing a conflict explanation.

In the default factual view, lines whose managed-memory qualification says answer_eligible: false are removed before evidence is shown to the answer model. If they are the only related memory, answer returns not_answered while preserving the related page identities; it does not call the model or claim the topic was never discussed. An inferred or explicit non-factual view may instead answer about a plan, report, historical decision, question, or discussion record while preserving that status. Report-only answers must keep the source attribution; an unattributed restatement is rejected before verification. For time-scoped factual items, current-value questions admit only intervals that are current at the reader clock.

An observation is offered to the answer model as one indivisible evidence item containing its readable L2 sentence and every current L1 leaf line. If any leaf cannot be re-read at its sealed fact, slug, line, and hash, the whole observation is withheld. A final observation citation expands back to every leaf source; the L2 page alone is never presented as sufficient proof.

Reranking is off by default for answer because generation and verification already select supporting evidence. --rerank opts into another sequential model call for especially noisy corpora or weaker answer models. --context returns the exact bounded evidence shown to the model for a review interface.

akno read products/zephyr-qx-100
akno read products/zephyr-qx-100 --from 10 --to 20
akno read --document doc_a1b2c3d4

akno list --kind tree --depth 2
akno list --kind pages --folder products

read does no ranking. It can return a source page in full even though ordinary recall quotes it narrowly. A document read reports whether the original, an indexed extraction, or only identity metadata is available. An exact page whose Markdown source is quarantined returns status: degraded with degraded: ["source_conflict"] and no page body. Repair the conflicting files and re-index; do not interpret that response as an empty or absent memory.

Use list before writing when the existing taxonomy or nearby page names matter.

akno graph --slug people/ada-marlow --hops 2
akno graph --query "Zephyr QX-100 warranty" --relation related_entity
akno graph --entity ent_01JEXAMPLE --direction out --history
akno graph --query "Ada Marlow plans" --memory-view planning

graph is not fuzzy search and does not answer questions. A slug or entity id gives an exact seed; a query extracts only declared exact names. Ambiguity stops traversal unless the optional contextual resolver already made a strongly supported select-or-abstain decision.

Traversal is bounded to three hops and 100 returned paths, with separate fan-out limits. Reaching a bound marks the result degraded so a partial graph never looks like proof that no other path exists. Paths contain compact node and relation identities plus source locators, not copied page claims. Follow the locator with read.

Valid level-one markers also project memory nodes and evidence-bound corrects, supersedes, contradicts, fulfills, answers, and caused_by edges. Graph traversal applies the same memory view as recall; duplicate or missing relation targets produce no edge. These rows are disposable projections of the marker and its exact payload, not a second memory store.

akno timeline --since 2031-01 --until 2031-06
akno timeline --match warranty
akno timeline --subject people/ada-marlow
akno timeline --scope today --timezone Europe/Amsterdam
akno timeline --view actionable --order nearest
akno timeline --source deadline --clock overdue
akno timeline --as-of 2031-04-12T10:00:00+02:00 --timezone Europe/Amsterdam
akno timeline --source document

Timeline combines three sources without collapsing their meanings:

• authored dated lines remain event results;
• typed retained time becomes event, state, plan, or deadline according to its relation;
• a dated orphan document remains document_evidence, never an event or canonical claim.

The default history view keeps inactive, rejected, completed, and cancelled records visible. The actionable view admits only active or accepted scheduled work and deadlines; it is a query, not a reminder or scheduler. Filter further with --status, --disposition, --source, an exact --clock relation, or the broad --scope past|today|future|all. --source both remains a compatibility alias for all.

Every result is classified at read time as past, today, current_period, ongoing, future, overdue, or undated. The response includes the exact as_of, IANA timezone, local date, and grouped counts, so a caller can repeat the same query deterministically. Month and year precision remain partial periods; Akno does not invent a day. Open validity intervals may be ongoing. Only an active or accepted due item can be overdue.

--since and --until use inclusive overlap semantics. Recurrence expands only when both bounds define a closed range and is bounded; an unbounded query returns only the stored anchor. A half-bounded range also returns the anchor when it overlaps, but reports degradation because later matching occurrences cannot be enumerated authoritatively. An instant recurrence needs an IANA timezone because a numeric offset cannot describe later daylight-saving transitions. If expansion or the retained temporal projection is incomplete, the result is degraded rather than falsely authoritative.

Owned documents remain grouped under their pages. For orphan evidence, file-created or file-modified metadata is used only when no supported date can be extracted, and is labelled as metadata rather than authored fact.

akno context "The Zephyr QX-100 needs service again" \
  --budget 8000 --pin products/zephyr-qx-100

The default context profile fits explicit pins, recent time evidence, a structure outline, and recall into one shared budget. It avoids four individually reasonable calls overflowing the host model when combined.

A host that checks memory before every substantive turn should use the precision-first profile:

const bundle = await memory.context({
  profile: 'auto_recall',
  query: currentUserPrompt,
  conversation_context: recentTurns,
  budget: 1200,
});

auto_recall is evidence preparation, not another answering model. It omits ambient pins, timelines, folder outlines, and generated summaries. Exact page lines and document quotations keep their locators.

A strong exact or semantic match can activate directly; borderline candidates need qualification. Exact evidence excludes semantic near-duplicates. Attribute questions require evidence for the attribute, not merely the entity name. Conflicting values cause abstention. Recent conversation can resolve a narrow local reference, including a possessive or short question fragment with an omitted subject, but cannot make unrelated memory relevant. Akno binds that reference to exactly one candidate identity before qualification; a qualifier may judge support for the requested field but cannot choose the conversational subject.

The content-free activation receipt reports reference_resolution as not_needed, resolved, ambiguous, or unresolved. The latter two return no automatic evidence. This lets a host distinguish ordinary low relevance from a follow-up that needs a clearer subject without logging the conversation or resolved identity.

Temporal qualification is applied before automatic activation. Current prompts cannot activate expired valid states; future prompts may activate asserted, active or accepted plans, while generic factual prompts cannot turn planned or scheduled memory into current fact.

Automatic context resolves the same semantic memory view from the current prompt before activation. This lets an explicit planning or report request inject the qualified record while an ambiguous factual prompt continues to exclude it. Hosts may pass memory_view when their own intent router already made that choice.

empty is a normal “inject nothing” result. The host should call the profile at most once per turn, place non-empty evidence inside a clearly delimited untrusted-memory section, and never persist the returned bundle as new memory.