Glossary

The documentation uses each of these terms in one sense only. Where a term has a field or value in the API, it is shown in code.

Products

RememberStack : The open-source memory engine for AI agents. It turns documents into claims, facts and entities, and serves them to agents without a language model on the read path.

remember package : The Python client, CLI and MCP server for RememberStack (pip install remember).

Self-hosted : Running the RememberStack engine yourself, with Docker Compose. See Install.

Deployments

Deployment : One running RememberStack memory: its own database, storage and credentials. Everything you ingest and query lives in exactly one deployment.

Sources and documents

Document (doc_id) : One logical file over its whole life, such as a spec or a meeting transcript. Its identity stays the same when its content changes. See Documents, versions and sources.

Version (version_id) : One snapshot of a document's bytes. Versions are append-only and never edited. See Documents, versions and sources.

Source (source_kind + source_ref) : Where a document comes from, given as a pair: the class of source and the item's stable ID within it. The pair is the document's identity. See Documents, versions and sources.

Content object : The stored bytes of a file, kept once per content hash however many versions or documents share them.

Representation (representation_id) : One conversion of a version into text (document.md). Evidence positions point into a representation.

Section : A part of a document found from its structure, such as its headings, with a role. Chunks and claims know which section they belong to.

Chunk (chunk_id) : A passage of a document's converted text: a run of whole blocks, not overlapping any other chunk. Chunks are what claims are extracted from and what passage search returns.

Versioning mode (snapshot, living) : What a new version of a document means. In snapshot mode (the default) every version stays standing testimony; in living mode the newest version replaces what older ones said. See Updating a source.

source_modified_at : When the source says the content was written or last changed. It becomes the said-on time of every claim from that version.

source_version_ref : An optional revision marker from the source system, such as an ETag or commit SHA, stored on the version.

Ingested by (ingested_by) : The actor that created a version: a user, an api_credential or a service. See Documents, versions and sources.

Claims and facts

Claim (claim_id) : One standalone statement a source made, tied to the exact text that supports it. Claims never change. See Claims.

Selection : The first extraction step: decides which statements in a chunk become claims and records a reason for each one it drops. See Claims.

Claimify : The second extraction step: rewrites kept statements into standalone claims and cites their supporting passages. See Claims.

Grounding gate : The deterministic check that refuses any claim whose cited passages or added words are not in the source. See Claims.

Attributed claim (is_attributed) : A claim that records someone's statement or stance ("Dana believes…") rather than asserting something directly.

Current testimony (is_current_testimony) : Whether a claim still counts as what its source currently says. Claims stop being current when re-extracted, superseded in a living document, or deleted. See Claims.

Fact (fact_id) : A piece of knowledge the memory holds true, built from claims: either a relation or an observation. Facts change as evidence arrives. See Facts.

Relation : A fact that connects two entities with a predicate ("Ravi works_on billing migration"). Relations are the edges of the graph.

Observation : A fact about one entity: a value, property, state or stance. Observations are searchable but are not graph edges.

Predicate : The name of a relation's connection, from a governed vocabulary of 16 core predicates, an other: escape value, or an extension pack. See Facts.

Extension pack : A named set of extra predicates for a domain, such as work.

Adjudication : The step that decides how a new assertion changes the facts: add, confirm, adjust, supersede or contradict. See Facts.

Evidence count (evidence_count) : The number of distinct documents whose current testimony supports a fact.

Support withdrawn (support: "withdrawn") : A fact whose only support disappeared because of a processing change rather than a source change. It is still returned, flagged. See Facts.

Retraction : A fact losing belief because a living source removed, or a deletion took away, its only support. Recorded as retracted_source_removal. See Updating a source.

Label : A fact's readable sentence, without dates.

Entities

Entity (entity_id) : One real-world referent (a person, team, project, event) however it is spelled. Entities have no type. See Entities.

Alias : A spelling a source used for an entity.

Profile : A short prose summary of an entity built from its important facts, used to decide identity.

Resolution : Matching a name to an entity. At write time a cascade of steps decides; at query time resolve_entity returns candidates and never guesses. See Entities.

Tier (T0–T3) : The resolution step that found an entity candidate: exact alias, similar spelling, similar sound, profile embedding.

Merge : Redirecting one entity into another when both are the same referent. Merges keep a snapshot and can be undone.

Evidence and disagreement

Evidence : The claims that support or dispute a fact, and the source text that supports a claim. See Evidence.

Evidence span (evidence_spans) : A character range in a version's converted text that supports a claim. A claim has one or more.

Stance (supports, contradicts) : Whether a claim backs a fact or disputes it.

Contradiction : Two or more facts that cannot all be true, returned together as a contradiction group. See Contradictions.

Co-member : Another fact in the same contradiction group.

Corroboration : Independent documents stating the same thing, counted by distinct documents, never by versions or repetitions.

Supersession : A later fact replacing an earlier one because the world changed; the earlier fact's window is closed at the later one's start.

Transcript : A fact's decision history: every adjudication, with outcome, method and confidence. See Evidence.

Time

World time (valid time) : When something was true or happened: valid_from, valid_until and valid_precision on facts; the claim_valid_* fields on claims. See Time.

Said-on time (asserted_at) : When a source made a statement.

Belief time (ingested_at, invalidated_at) : When the memory learned a fact and when it stopped believing it.

Precision (valid_precision) : How exact a window is: instant, day, month, quarter, year, open (known start, ongoing) or unknown.

Claim valid kind (claim_valid_kind) : What a claim's stated time describes: event_time, effective_period, measurement_period or proposition_validity.

Time mode (current, at, overlap, history) : Which facts a query selects by world time. See Time.

Temporal match (confirmed, possible) : Whether a fact's window establishes that it matches the query's time, or only fails to rule it out.

Temporal scope (temporal_scope) : The part of every result that states which time it describes and when it was evaluated.

Processing

Pipeline : The stages that turn an ingested version into claims, facts and search indexes, in the background. See The pipeline.

Stage : One step of the pipeline, such as convert or extract_claims, with a status per version.

Dead letter (dead_letter) : Work that failed and will not be retried automatically.

Readiness : Whether given versions have finished processing and the capabilities you need (pipeline, p1, live_graph, p3) are ready, so the content can be recalled. See The pipeline.

Reading

Assured operation : One of four fixed, registered reads: resolve_entity, facts_context, claims_and_sources_context, combined_context. See Retrieval and Assured operations.

Primitive : A lower-level read such as search, lookup, graph traversal or hydration. See Retrieval.

Envelope : The object every read returns: the results plus an account of their grain, time, completeness and emptiness. See Reading a result.

ContextBundle/v2 : The result of combined_context: a claims-and-sources envelope and a facts envelope side by side.

Grain (fact, evidence, composite, compiled) : What kind of truth an envelope holds.

Negative (unknown_entity, known_empty, boundary) : The typed reason an answer is empty.

Hydration : Re-reading search candidates from live data before returning them, and dropping those that no longer hold (dropped_by_hydration).

Semantic search : Search by meaning over embeddings (vectors), with pgvector.

BM25 : Keyword search that ranks by term matches, with PostgreSQL's pg_textsearch.

RRF (reciprocal rank fusion) : Merging several ranked lists by summing 1 / (60 + rank) for each item.

Query space (memory_v1) : The prepared, read-only views and functions that SQL queries run over. See Query space.

SQL queries : Reading memory with SQL over the query space. Every statement is parsed and validated against the query space before it runs. See Explore memory with SQL.

Saved query : A reviewed SQL query stored under a name and run with parameters. See Saved queries.

MCP : The Model Context Protocol, which lets coding agents call memory as tools. See Connect your coding agent.