Skip to main content
Meerkat Recall’s Living Wiki is the human-readable semantic projection of canonical knowledge. It does not create evidence and its output must never be re-ingested as a source. A source record, an entity row, and a compiled projection are not automatically wiki articles. Publication requires editorial selection, semantic organization, and statement-local support. Use this page to understand the projection and publication contract. To try it, jump to the runnable demo. To call it from an agent, use the MCP surface and inspect the connected server’s tools/list schema. The implementation has three deliberately separate surfaces:
  1. Wiki answers “what does Meerkat Recall know?” with curated articles.
  2. Evidence answers “why does Meerkat Recall believe this?” with source content, quotations, dates, and provenance.
  3. Inspector answers “what did Meerkat Recall’s machinery do?” with record IDs, revision IDs, hashes, compiler receipts, dependency inventories, and ingest state.
The graph is canonical memory. The wiki is one disposable, rebuildable view of it. Technical metadata must not become the article merely because it is easy to render. Surface and editorial state are explicit stored fields, not inferred from a record’s shape. Entity and graph-selection compilers assign inspector, and source compilers assign evidence. Only the distinct article kind may be assigned to wiki, and only an LLM editorial compiler may create that kind. The renderer additionally requires explicit published state. Topic, community, and query projections cannot become wiki articles by changing a flag.

​
Research basis and Meerkat Recall adaptation

Karpathy’s original LLM Wiki idea file separates immutable raw sources, an LLM-maintained wiki, and an evolving schema. Its central useful move is compilation: integrate each source into durable, interlinked pages so later questions do not reconstruct the same synthesis from raw chunks. It also calls out ingest, query, and lint loops, filing useful query results back into the wiki, contradiction detection, source references, and a human-browsable graph. The later Retrieval as Reasoning paper sharpens the pattern into three properties: compilability, composability, and evolvability. It adds progressive link traversal plus a persistent error book, and reports that structural validation, source-grounded verification, and cross-page consistency checks are all needed because unsupported facts, malformed references, dangling links, and contradictions recur in generated wikis. Meerkat Recall adopts the compounding read model but changes the authority boundary:
  • the canonical graph and immutable source evidence remain authoritative; wiki Markdown is a rendering and can never become evidence by being re-ingested
  • graph edges replace fragile text-only backlinks, while stable views and immutable revisions provide human and agent-readable pages
  • dependency and selector-trigger indexes turn graph/source changes into dirty projections without scanning every page
  • truth slots, claim freshness, source watches, conflicts, and validation receipts make lint and maintenance continuous rather than occasional
  • MCP exposes direct page reads and compositional graph traversal to agents
  • the required LLM editorial compiler must pass statement-local evidence entailment; fluent uncited prose is a publication failure

​
Wiki article model

Everything with a durable identity can exist as a graph entity, but existence does not imply article prominence. The article planner must select subjects using context-dependent relevance, importance, graph centrality, recency, and human editorial guidance. Peripheral entities can have very short pages whose main purpose is to explain why they occur in this particular knowledge space. An LLM writes each page from a bounded editorial context containing graph facts, narrative observations, first-class events, current truth, relevant history, and source excerpts. It selects stable semantic sections according to the subject and available knowledge. A person might have Family, School, Work, Interests, Current matters, and Timeline. A device might have Owner, Configuration, Integrations, Problems, and History. These are not database-table dumps. The LLM’s editorial job is to synthesize, choose, and organize the knowledge that helps a reader understand the subject. A deterministic traversal can prepare its context, but cannot supply article prose. Events remain first-class graph records and supply chronology. Current truth is presented before superseded history. Conflicts become plain-language notices with evidence links. Source documents normally support articles rather than becoming articles themselves.

​
Authority boundary

The authority chain is one-way: source -> doc revision -> evidence span -> assertion -> graph and truth state -> knowledge view revision A stable knowledge_view record provides identity, security scope, current state, and a pointer to an immutable knowledge_view_revision. Each revision contains:
  • structured blocks and rendered Markdown
  • a statement-level support ledger
  • citations to graph records, assertions, sources, and immutable evidence spans
  • the complete dependency inventory used by the compiler
  • claim and projection freshness metadata
  • deterministic validation and compiler receipts
  • a pointer to the superseded revision

​
No filler prose

Factual prose is not an unstructured model completion. Every claim-bearing passage is a KnowledgeStatement with:
  • the exact text published in the page
  • a derivation mode: direct, synthesized, or verbatim quote
  • citation IDs supporting that exact statement
  • a verification mode: deterministic typed-record rendering or evidence entailment verification
The factual block Markdown must be exactly reproducible from its statement ledger. Publication fails when a factual block contains text outside that ledger, a statement has no citations, a citation is not a declared dependency, or a verbatim quotation lacks an immutable evidence span. This makes “citations nearby” insufficient: the support contract is statement-local. The deterministic compilers emit no boilerplate for missing summaries. Sparse knowledge therefore remains terse instead of receiving invented or vacancy-filling prose. Source evidence can add bounded verbatim excerpts from selected immutable revisions, with exact byte offsets and quote hashes, but those excerpts belong on the evidence surface. A model-based editorial compiler must pass the evidence-entailment gate before its prose can be published. Model fluency alone is never a publication signal. The editorial compiler also rejects type-only leads and implementation-facing filler such as claims that a subject is central to an editorial context, inventory, or graph. Rejected drafts enter the same bounded LLM repair loop as unsupported claims. A page must communicate concrete supported knowledge, not merely produce grammatical text. Identity metadata alone is not enough to publish an article. The lead must cite a concrete canonical summary, governed relationship, event, assertion, or source passage. Candidates with neither a summary nor a governed connection are excluded before planning, and evidence-scarcity boilerplate is rejected rather than presented as an article. Generated prose is disposable. It is never canonical memory, never feeds back as evidence, and is regenerated when its dependency or quality receipt changes.

​
Graph quality as a publication gate

A living wiki makes graph defects directly visible to people. Publication must therefore consume categorized graph-quality findings, not reduce them to a single opaque score. At minimum the receipt must represent:
  • non-durable or disconnected entities
  • duplicate labels and possible duplicate identities
  • alias hygiene findings
  • missing, uncontrolled, inactive, self-referential, or otherwise invalid relationship predicates
  • unsupported, stale, superseded, or conflicting claims
Each finding has an affected-record set and a publication action: omit the claim, keep the article in draft, or publish a semantic warning. Possible duplicate people are article-blocking because a biography can otherwise merge two people. Invalid predicate edges are claim-blocking because rendering them turns graph noise into authoritative-looking prose. Application-specific graph audits may supply additional classifications. Meerkat Recall owns the generic typed issue and publication-receipt mechanism; application policy must not become untyped canonical graph truth.

​
Freshness and change propagation

Canonical event, relationship, and attribute promotion now creates or updates one deterministic freshness record per claim. View revisions summarize the minimum claim score, stale truth slots, unresolved conflicts, unsupported claims, expired claim reviews, and the next review deadline. A passed review deadline degrades the projection even when no graph record changed. Publication atomically materializes one current reverse-dependency edge per revision dependency. Outbox changes use the (space_id, dependency_record) index to resolve only affected views, then move each stable view to dirty. The immutable revision retains the historical full dependency receipt while the reverse index is replaced on every publication. The next normal read can recompile a dirty view. Reads choose one of three policies:
  • allow_stale: return an existing revision without rebuilding
  • prefer_fresh: rebuild when needed, but return the last authorized revision with a warning if refresh fails
  • required: fail rather than return a revision that could not be refreshed
Forced maintenance rebuilds always publish a new immutable revision. Unchanged reads take the hot path: a current revision with the active compiler version and an unexpired freshness receipt returns before graph traversal. The reverse-dependency index is therefore the cache-coherence authority; canonical changes mark only their dependent views dirty, and dirty views cannot use this fast path. Web sources can additionally carry a source_watch. It holds the mutable schedule, HTTP validators, last observed hash, retry state, and next deadline. Every attempt appends an immutable source_watch_check receipt with the final validated URL, validators, content hash, byte count, outcome, and any bounded error. The maintenance timer only enqueues due checks; network retrieval runs as a leased refresh_source_watch work item with normal work-run and retry visibility. Remote retrieval accepts only HTTP and HTTPS, rejects embedded credentials and non-public destinations, revalidates every bounded redirect, pins validated DNS resolution into the request client, enforces timeouts, and streams through an explicit size limit. A textual change updates the existing source and document through the source-first re-ingest contract, marks dependent views dirty, and reopens canonical triage. ETag/Last-Modified and hash-unchanged observations do not create spurious source revisions. Non-text media and non-web connectors are recorded as unsupported until their source-specific refresh adapters exist.

​
Security

Compilation is exact-envelope, not union-envelope. The root entity’s complete security object is fingerprinted into the stable view key. Only dependencies with that exact security envelope can contribute to the revision. MCP checks authorization before compilation and again on the resulting view and revision. This prevents a low-clearance projection from leaking the existence or labels of higher-clearance records.

​
MCP surface

  • get_knowledge_view accepts exactly one entity_id, source_id, or definition_id and is available in the Default, Agent, Maintenance, and Full profiles.
  • rebuild_knowledge_view is available in Maintenance and Full profiles and requires write authorization for the selected root.
  • materialize_knowledge_views scans bounded, resumable pages of existing entities, sources, and definitions. It applies record-level ABAC, isolates per-root failures, returns an independent cursor for each root kind, and writes only derived projection records. It does not promote raw entity or source projections into the wiki. Within a root kind it uses explicitly bounded concurrency over graph-read indexes, so graph-scale databases do not turn one MCP call into an unbounded serial traversal.
  • get_knowledge_view_definition reads the governed typed selector.
  • upsert_knowledge_view_definition creates or updates a bounded topic, seeded-community, or safe saved-query selector and its explicit draft/published/hidden context-plan state. published makes a context eligible for the LLM editorial compiler; it does not expose the deterministic projection in the wiki. Raw database queries are not accepted.
  • compile_wiki_article runs a Meerkat-backed structured-output writer over the bounded context, then runs a separate structured-output verifier over every statement and its citations. The verifier independently requires both evidence entailment and relevance to the named article subject, so an off-topic fact cannot pass merely because shared source material supports it. Drafts and publications are immutable, dependency-hashed article revisions. Publishing requires an approved context plan and a clean freshness receipt.
  • get_source_watch reads the schedule and immutable checks without network work. It is available in every read-capable profile.
  • configure_source_watch and check_source_watch require source write authorization and are available in Maintenance and Full profiles. Manual checks enqueue the same leased, idempotent work item as scheduled checks.
  • get_entity_context_bundle exposes attributes, truth slots, events, assertions, and freshness in addition to graph relationships, with ABAC and time-window filtering applied to each record family.
The authenticated REST surface exposes GET /v1/knowledge-views and GET /v1/knowledge-views/{id} over all stored projections. The human surfaces are separate:
  • GET /wiki and GET /wiki/{id} expose only article-eligible semantic views
  • GET /evidence and GET /evidence/{id} expose source evidence
  • GET /inspect/knowledge-views and its detail route expose every projection and technical receipt
All three apply record-level ABAC, escape stored content, disable browser caching, and share the same canonical authority path. Evidence detail reads also run the idempotent source compiler, so a stale compiler receipt is refreshed before the source is rendered. Wiki reads never synthesize or publish an article implicitly.

​
Runnable projection demo

Run a fully isolated example using Meerkat Recall’s normal Meerkat LLM runtime. Copy deploy/native/demo-wiki-processing.example.json to a private configuration file, replace its account, model and immutable version values, and set the environment variable named by its credential_ref. The file must explicitly approve the isolated default space’s wiki_planner, wiki_writer and wiki_verifier bindings. In a native archive, the example is under service/. Then pass the configured file:
The command loads this architecture document as a real source, creates an immutable document revision, derives assertions with byte-addressed evidence spans, promotes supported relationships and an attribute into a small graph, compiles entity, source, and topic context projections, asks an LLM curator to choose a small interlinked article set, asks the configured writer LLM to write each semantic article, independently verifies every article statement, and starts the normal routes in local development mode. Open http://localhost:3000/wiki for the resulting article, http://localhost:3000/inspect/knowledge-views for deterministic projections and http://localhost:3000/evidence for source material. Model selection and provider authority come from the explicit processing configuration; credentials alone do not approve those operations. The default build requires that configuration and the selected provider’s credentials. A --no-default-features build needs neither and still demonstrates the evidence and inspector surfaces, but deliberately leaves /wiki empty:
Use --port 3100 when port 3000 is in use. The demo is an isolated in-memory store and disappears when the process stops.

​
Graph-derived editorial contexts

Topic, community, and saved-query context projections are rooted in a first-class knowledge_view_definition. Topic and community selectors traverse only approved relationships within explicit hop and member bounds. Saved queries use a typed entity filter AST with optional type, name/alias, and exact approved-attribute predicates. Every selected member states why it matched and cites the member plus the relationship or attribute claims used for selection. Current dependencies include table-trigger keys in addition to concrete record IDs. Outbox invalidation adds the corresponding key for each changed record, so a newly created relationship, entity, or attribute can dirty a projection even though that new record could not have appeared in its previous revision. Explicit member or scan bounds are disclosed in the revision and degrade its freshness state when they truncate selection.

​
Article planning and managed publication

Meerkat Recall has an explicit LLM article planner, LLM information architect, article writer, and statement-level verifier. create_wiki_from_graph scans the complete caller-authorized semantic graph with no configurable prefix or record ceiling. It isolates exact security envelopes and reads entities, relationships, assertions, attributes, first-class events, truth slots, and predicate governance to exhaustion. Every eligible entity is assigned exactly once to a bounded graph-neighborhood planning shard of at most 30 entities. The old global top-60 reduction is not part of the active path. The LLM planner must account for every entity, normally with one article per semantic subject. With the default max_articles: 0, the automatic budget is the complete eligible population, bounded at 2,000. An entity may be synthesized into another page only when Meerkat Recall identified the two records as a possible person-identity duplicate and the planner explicitly records the consolidation. The full plan must retain at least 70 percent as many pages as eligible subjects. A separate LLM organizer sees the complete planned corpus, infers the space purpose, selects the defining featured pages, and designs semantic reader lenses. Exhaustive placement is then performed by bounded 30-article LLM assignment calls against that fixed taxonomy. Meerkat Recall validates that every selected article appears exactly once before compiling any page. These are disposable navigation, not graph assertions or a pre-created folder tree. Meerkat Recall fills omitted article links from governed direct graph edges and resolved source-evidence co-mentions among selected subjects. It creates topic definitions for single subjects and seeded-community definitions for an explicitly consolidated identity cluster. Possible duplicates are never silently deleted. Every reader link must pass graph-or-evidence validation; shared security subjects alone never create navigation. The result reports complete scan counts, bounded planning partitions, possible duplicate-person groups, relationship and attribute rows withheld by predicate governance, eligible and planner candidate counts, selected and verified articles, failures, and hidden articles. Entity planning also carries earliest and latest semantic evidence times, so decade-old source volume does not masquerade as present relevance.

​
Recurring publication

Published creation is managed by default. It durably enqueues reconcile_living_wiki. A periodic run always performs the complete bounded scan, but a stable graph fingerprint reuses the current wiki without any LLM calls. Entity, relationship, assertion, attribute, event, predicate, truth-slot, or merge outbox changes debounce an earlier reconciliation and force it past the periodic cache decision. Historical outbox replay from before the managed chain is ignored. Older queued runs are cancelled, running work is not interrupted, and only one wiki reconciliation executes per space. A changed run replans, recompiles, hides pages no longer successfully retained, preserves immutable history, and schedules its successor. If any selected article fails, the fingerprint is not cached and a later run retries. No entity IDs, article definitions, navigation sections, or prose are seeded by the command or worker. Article compilation is independently bounded by ELEPHANT_WIKI_ARTICLE_CONCURRENCY from 1 through 32. Writer and verifier transport failures use bounded exponential backoff, while unsupported prose still fails closed. This permits large graphs to use provider and host capacity without weakening the evidence or verification contract.

​
Evidence available to the writer

For each planned entity, the editorial compiler also retrieves a bounded set of resolved assertions that mention it and validates their immutable evidence spans. These source passages are available to the writer and verifier even when the extracted predicate has not passed graph governance. They remain source evidence rather than canonical relationships, and their assertion, revision, and source records become explicit article dependencies. This lets a wiki synthesize real document content without laundering predicate-quality defects into graph facts. The verifier evaluates relevance as a private-wiki contract, not a general encyclopedia contract. A supported family trip, preference, purchase, message, decision, project, or event can fully explain why a peripheral place, product, organization, or concept occurs in this knowledge space. Sparse evidence may produce one short, concrete local-relevance fact; the verifier must not demand invented background or synthesis that the catalog cannot support. Entailment, temporal framing, citation resolution, and rejection of unrelated co-mentions remain fail-closed. Resolved assertions also provide bounded entity co-mention counts to the planner and the automatic navigation augmenter. These may create related-article navigation between selected subjects, but they do not assert a relationship type. The plan validator accepts a link only when a governed graph edge or resolved source-evidence co-mention supports that navigation. When a legacy assertion’s quote hash no longer matches, the compiler validates the recorded byte offsets against the immutable revision, derives a new citation span hash from those exact bytes, and records a warning. It does not mutate or trust the stale assertion hash. Legacy inline-promotion assertions can carry an empty assertion-level span and a derived: hash while linking to the real doc_node. For those records, the compiler resolves the node, verifies that it belongs to the same revision and policy, uses the node’s revision byte span, and records both the node and revision as dependencies. This is the supported bridge from derived assertions back to quotable source text.

​
Publication security and governance

Source evidence may carry a stricter envelope than the semantic entity. The article inherits the restrictive join: maximum sensitivity level, unioned labels and handling caveats, and unioned subject scope. Evidence under a different named policy schema is not mixed. A semantic record under the placeholder default policy may be promoted into the first named source policy; after promotion, all further named evidence must use that same policy. This prevents source-backed synthesis from laundering protected material into a less restricted entity page while allowing a named source policy to govern pages for older default-policy entities. Planning includes only relationships and attributes whose registered predicates are active, have semantics, and are in explicit Configured governance state. Review-pending predicates remain in canonical memory and quality workflows but cannot enter the LLM editorial context. Article neighborhoods default to one graph hop and permit up to 250 governed members; both bounds are explicit MCP inputs so a highly connected entity does not silently publish a truncated view.

​
Current limits

Only that compiler can create article projections for /wiki; deterministic graph projections remain editorial inputs under the inspector surface. The following remain implementation gaps, not implicit features:
  • connector-specific refresh adapters beyond textual HTTP/HTTPS web sources
  • a first-class persisted interchange record for external domain-specific graph-quality audits; active generic publication gating already withholds likely duplicate people and non-configured predicates and reports categorized counts, but application-specific audit taxonomies are not canonical Meerkat Recall graph truth
  • human approval queues for plans in deployments that require approval beyond the explicit publish request
  • article-local inline entity links inside prose; first-class related-article navigation is already stored on definitions and rendered only for published, authorized target articles
  • arbitrary model-selected quote retrieval beyond the exact evidence spans in the bounded context; those supplied quote bytes are hash-verified before the writer or verifier sees them
  • predicate-specific freshness policy configuration and backfill