The lineage-first fabric and the engine underneath it

Each agent owns a local memory graph. There are no raw cross-graph edges; exchange occurs only through typed artifacts validated at the boundary (ClaimEnvelope, RetrievalTrace, EvidenceExport), never through raw graph reads or a global merged store.13 A claim is supported by authorized derivations, and knowledge need not be stored as a literal sentence: independent holders can jointly reconstruct an answer no single one of them holds.4 An artifact carries owner, scope, temporal validity, lineage and authorization; a derivation records support, not a guarantee of truth.5 Reconstruction is demonstrated in the deterministic fixture and the Gate-1 tests, not in a deployment.6

Formally, agent i owns G_i = (V_i, E_i, ψ_i); a materialized artifact is a = (id, owner, z, Σ, τ, λ, p, refs): semantic state, scope, temporal validity, lineage, authorization and opaque evidence pointers; it is an optimisation and policy boundary.7

The formal definitions of the local graph and of a materialized artifact are held in the manuscript under review and will be added here after the review decision.

A question is a first-class, bounded, routable request rather than control flow: it carries an asker, a trigger, a scope, a budget, target units and a status, and it may travel when private memory cannot.8 Responses preserve contributors, evidence roots and unresolved disagreement; disclosure still requires authorization, because even a question can expose sensitive intent.9 In the organisational benchmark the artifact is implemented as a request for order-(k+1) cell counts answered by one-cell sketches.9

Formally, Q = (id, asker, q, trigger, Σ, τ, u, λ_ctx, D, π, B, status): an uncertainty, a motivating lineage, candidate domains, a policy and a budget.10

The formal tuple definition of the QuestionArtifact is held in the manuscript under review and will be added here after the review decision.

A horizontal plane coordinates peers; a vertical axis forms higher-level representations (in the benchmark, worker → team → department → region → executive; in the engine, MESSAGE → EPISODE → TOPIC → PERSONA); lineage threads preserve the evidence and transformations supporting each artifact and form a derivation DAG.11 Cross-functional projects remain overlapping scopes rather than forced branches of an organisation tree. Hierarchy is prior art; what is specific here is that the vertical axis never severs the lineage thread.12

Support is counted by distinct independent evidence roots after a correlation discount, not by replica count: exact copies count 0; the same worker counts 0 beyond the first; same team 0.5; same department 0.8; otherwise 1.0.13 A verifier is not shown the current answer where that would contaminate independence; disagreement becomes a conflict object, never an automatic merge.14 Typed lineage retains the difference between a shared database and a shared human source; active lineage diversification chooses the acquisition action with the largest expected gain in survivability margin per cost.15 Two further design points, a lineage typing that keeps a shared database distinct from a shared human source and an active choice of which evidence to acquire next, are described in the manuscript under review.

The five epistemic functions are verification, contradiction, relation discovery, hypothesis generation and prediction/falsification; they are architectural functions, not all measured natural-language capabilities.they are roles the architecture defines, and not every one of them has been measured as a natural-language capability.18

A graph-native memory and retrieval system for long-running AI agents.2 It turns conversations and events into structured, searchable memory; instead of a flat vector store it combines semantic retrieval, graph relationships, speaker identity, temporal reasoning, reranking and evidence attribution to recover the right context for an answer.21 Most agent-memory systems are good at finding text that looks similar to a query and weaker at questions that require identity, chronology, aggregation or multi-step reasoning (a positioning statement, not a measured comparison).22 Six question classes the engine targets: who said what; when did it happen; what changed over time; what facts belong together; why was this answer produced; what should happen when memory is insufficient.23

Memories are nodes with metadata and relationships rather than isolated text chunks, preserving identity, sequence, provenance and semantic connections.24 Nodes live in a MESSAGE → EPISODE → TOPIC → PERSONA hierarchy with a TEMPORAL anchor layer and are linked by six typed, weighted, signed edge types: TEMPORAL, SEMANTIC, ENTITY, CAUSAL, HIERARCHY and CO_ACTIVATION.25 Time and speaker identity are first-class metadata; relative expressions such as “yesterday” and “last week” are resolved against the original message timestamp.26

There is an index. Stage 1 is hybrid recall over vector similarity (a per-session LSH index with brute-force fallback for small sessions) and keyword search; later stages expand by entity edges, temporal chains, hierarchy traversal and co-activation; a final reranking stage scores candidates with several signals: semantic similarity, keyword match, entity overlap, temporal relevance, speaker-aware boosts, specificity bonuses and reranking deltas (signal names only; the design document’s weights are boost multipliers, not a normalised split, and are not shown).27 One reranking pass is an interference-style sweep over time × hierarchy, the mechanism the previous site called “wave-based resonance retrieval”; it is one signal in a hybrid pipeline, not the retrieval method, no isolated measurement attributes gains to it, and the measured gains on LoCoMo came from per-agent routing and dialogue-pair back-fill.28 Queries are routed into strict fact extraction, temporal, list, aggregation, memory-grounded inference or open-domain fallback modes.29

The router just described (query-type detection, per-store retrieval, weighted fusion) is the code artefact the company also describes in a patent filing, reproduced below as supplied.70 This is the retrieval Tesseract, not the coordination Tesseract of Part 1 (see “Two disambiguations” below); the two share a name for historical reasons only.68

The engine records what happened at each retrieval stage, including candidate counts, latency, score components, reranking effects and the evidence supplied to the final answer, so every retrieval failure is measurable.30 This is the engine-level counterpart of lineage in the fabric.

Retrieval writes back: co-activation edges strengthen when nodes are retrieved together (capped at 3×) and timing-dependent plasticity adjusts temporal edges.31 A consolidation cycle decays activation, prunes weak edges, merges near-duplicates above a similarity threshold, creates hub summaries, promotes hot nodes up the hierarchy and evicts cold ones.32 Stated plainly so that operators know: retention is automatic unless configured, and there is no user-facing purge-on-demand feature today.33

The answering and embedding pipeline runs against a local model server: any OpenAI-compatible endpoint such as LM Studio, or Ollama, chosen from the configured base URL.34 The September 2026 campaign used gemma-4-e4b with nomic-embed-text v1.5.35 Model names are configuration, not product facts, and this site states no default model because the README and the code differ.36 Disclosure: the engine includes an optional open-domain fallback retriever for questions memory cannot answer; whether it is enabled is a configuration decision a deployment must make explicitly.37

Persistent SQLite or in-memory storage with indexes partitioned by session key and one LSH index per session.38 Session isolation is data scoping inside one storage layer; it is not encryption, tenancy or a security boundary, and this site does not present it as one. Multi-agent namespaces and permissions are a roadmap item.39

Under active development; current work focuses on single-hop extraction, temporal questions, list and aggregation queries, query routing, latency tracking and benchmark attribution.40

Planned, not done.39

• planned publish reproducible benchmark results
• planned a minimal quick-start
• planned a documented API for the memory graph
• planned graph visualisation tools
• planned additional local and hosted model providers
• planned multi-agent memory namespaces and permissions
• planned a packaged library

Consequently there is no pip package, Docker image, SDK or REST API today. The source repository is public to read at github.com/anovruzov/NeuralGraph, but the licence has not been selected, so all rights are reserved until one is and nothing in it is offered for use or redistribution.4169

Intended use cases (not deployments): persistent memory for autonomous agents; personal assistants that remember conversations accurately; multi-agent shared memory; timeline and event reconstruction; long-horizon research assistants; benchmarkable retrieval-augmented generation; explainable memory retrieval.42

Rule: a word without a mechanism behind it is not used on this site. Spores, resonance and waves are out; the table says what each remaining word means in code or in the benchmark.48

• Two ladders. manuscript-B0 to B7, B3Q, B6C (Class D) versus bench-B0 to B9 plus ORACLE. “B7” means lineage fabric plus questioning in the manuscript and lineage-aware hierarchy without questioning in the benchmark; the site always prefixes.67
• Two Tesseracts. Tesseract (coordination) is the cross-node coordinator with typed contracts; Tesseract (retrieval) is an older single-node query-routing module with temporal, entity, reasoning and adversarial stores; they share a name for historical reasons only, and this site means the coordinator unless it says otherwise.68

Evidence chips.

• A measured retrieval
• B deterministic fixture
• C synthetic acquisition simulation
• D population-scale symbolic simulation
• T deterministic coordination tests
• IP in progress, no results
• W withdrawn

Status pills.

• measured
• simulated
• implemented
• tested
• configuration
• design
• proposed (hatched)
• in progress
• superseded
• planned