Concepts

RememberStack has two deliberately separate truth layers. Claims are immutable source testimony (“what was asserted, by whom, when”); facts—relations and observations—are the adjudicated worldview (“what the system holds or held true”): supersession-adjudicated, clocked on two time axes (when a fact held in the world, and when the system learned it), evidence-counted per distinct source—repetition is not corroboration—and contradiction-tracked. The fact_claim_evidence association is the auditable bridge between the layers, recording which claims support or contradict each fact. Query claims to inspect testimony; query facts to answer current or historical truth questions, then follow the bridge to see why the system believes or believed the fact.

(Internally these guarantees are decisions D41 and D54.)

One distinction carries the whole system: what a source said is not the same as what is currently true. Sources go stale, contradict each other, and get corrected. Memory that conflates testimony with truth either serves outdated statements as facts or silently rewrites history.

This page is the vocabulary every other page assumes.

TL;DR for agents

Default motion: orient → verify → audit.

Terminology ladder

Example: “the roster says Bob is CFO” enters as a claim, adjudicates into a relation (Bob works_for Acme or a role relation), and may appear in a compiled K page. A principle like “we never promote on rumor alone” is authored knowledge that may cite that history — it is never inferred from it.

Grains — typed answers

Every assured response is labeled with the grain of what it contains:

When an agent needs both testimony and facts, combined_context returns ContextBundle/v2 with complete claims_and_sources and facts envelopes. It is not a composite envelope and does not blend or re-rank the two authorities.

Contract: current-fact answers may be assembled only from validity-filtered facts. Claim asserted-validity intervals are testimony, never verdict. The registry linter rejects operations that would smuggle claims into current_facts intent.

Evidence-grain defaults to current testimony: re-extracted or living-version-superseded claims are excluded unless you opt into historical/superseded modes. See Lifecycle.

Two clocks

Facts carry two independent axes:

Compose multi-step temporal questions from ordinary calls: resolve an event time from one result, feed it as valid_at to the next. No special “timeline engine” is required.

Media locator is a third coordinate (offset in a recording / region in an image). Do not mix it with validity or belief time.

Supersession, not overwrite

New information closes a validity window on the old fact. Nothing is silently erased. You can still ask what was believed before the close.

Contradictions are loud

When sources disagree, both positions stay. They form a contradiction group. Retrieval must return co-members (bounded inline + continuation). Returning one side alone is a contract violation, not a ranking preference.

Evidence counting

Support count = distinct document lineages with current testimony supporting the fact.

Does not inflate from:

• re-extraction with a new model generation
• version churn on one lineage
• repeated sentences inside one document

When sole current support disappears:

• Living document removal → fact closed (retracted_source_removal) — recorded, reversible
• Toolchain re-extraction that fails to re-derive → support may show withdrawn and needs caveat + audit (not silent vanish)

Identity

Entities live in a canonical registry with aliases and tiered resolution.

An entity profile is orientation, not identity or a type. It is a disposable projection of current supported observations and relations. Write-time T3 compares the mention name plus its claim context with the candidate name plus profile and salient facts; it never compares two name-only vectors. An entity with no evidence has no T3 vector and reaches T4 instead. T4 receives every candidate in the bounded snapshot with aliases, current profile description, salient statements, and its T3 score or gate.

At write time, an exact name is normally only a T0 candidate signal. It never silently accepts or merges an entity across documents. There is one bounded exception: after T4 has matched that exact canonical name to one entity in a document, later exact repeats in the same document may replay that match at T0. The replay is disabled if the binding is absent, ambiguous, stale, or not built at the current document-t0-v1 generation; fuzzy T1/T2 candidates never gain this authority. One candidate with a current profile may otherwise be accepted by T3. Empty or conflicting evidence, and every multiple-candidate case, reaches one binary, match-biased T4 call on the configured simple model. It selects one supplied candidate id or new. Compatible evidence favors the existing entity; missing overlap and different topics are not evidence of a different referent. new requires positive distinction from every supplied candidate. Confidence remains in the audit but never routes to a frontier model, and candidate truncation remains visible without creating a third decision state.

Profile refreshes are evidence-change debounced. If evidence is withdrawn, the cached summary and vector rebuild or clear. Hard-forgetting one source of a shared entity rebuilds from the surviving sources, so forgotten phrases do not remain in the summary or semantic vector.

A successful evidence-backed profile publication nominates that entity's alias neighborhood for bounded convergence. Automatic merging remains off until its profile-space threshold is calibrated; equivalent open review proposals are deduplicated, and every accepted merge remains a reversible redirect.

Merges follow survivor chains with redirects disclosed. For nickname/misspelling recall at read time, use evidence search (claims_and_sources_context) rather than assuming fuzzy resolve. There is no deployment-wide exact-name auto-accept switch in the shipped cascade.

Document identity (write path)

Three IDs stay distinct:

Versioning mode per lineage: snapshot (archival successive reports) vs living (standing statement / wiki). Details: Lifecycle.

Progressive disclosure of knowledge (Plane K)

K pages are prose syntheses over the fact layer:

• Compiled pages regenerate when routing rules say evidence changed.
• Authored pages (plans, principles) are never machine-rewritten; evidence drift raises flags.

K is orientation. Facts are verification. Claims are audit.

Next

• Architecture — planes, stores, end-to-end flow
• Ingestion — how claims and facts are produced
• Retrieval — how agents consume them