Version. 0.1
Date. 2026-07-27
Status. Investigation. Field scan of current graph-memory work for artificial-intelligence agents, mapped against Loomworks commitments and build state.
Author. Claude.ai (investigation layer). Operator: Marvin Percival.
Provenance. Produced in the session that also produced loomworks-graphs-explained-v0_1 (operator explanation) from loomworks-graph-role-elaboration-v0_1 (Claude Code read-only engine inspection, loomworks-record commit 3321fa6). The field material comes from web sources current to 2026-07-27, listed in the Sources section. Engine-state claims below repeat the elaboration document's findings; they were not re-verified in this session.
The artificial-intelligence agent industry has recently converged on graph-structured memory — systems that store what an agent knows as connected, time-aware statements rather than piles of text. This document scans that field and maps it against Loomworks. The findings: the field independently validates the substrate Loomworks committed to at the seed; it sharpens the priority of two pieces of already-queued work (the forward-walk query and the memory-equals-recall reframe); it contributes three concrete, adoptable ideas (hybrid retrieval, as-of-a-date queries, supersession-aware retrieval scoring); and it presents one pattern that must be refused (autonomous graph-building with no human commit gate). It also hands Loomworks a positioning line: governed graph memory versus autonomous graph memory.
Decisions this document feeds, none required immediately: the memory-equals-recall reframe scoping (hybrid retrieval belongs in it), the forward-walk build (validated as the valuable thing), and prospect-document language (the positioning paragraph in section 6).
The production systems now leading the agent-memory category — Zep (with its Graphiti engine), Mem0, Letta, Supermemory — and the current research strand (MAGMA, GAM, AriGraph, temporal semantic memory work) share a converged shape:
Memory as a temporal knowledge graph. Facts are stored as connected statements with validity windows — when a fact became true, and when, if ever, it was superseded. Superseded facts are not deleted; they close their window and remain. Entities evolve over time. Every derived fact traces back to the raw episode that produced it.
Hybrid retrieval. Recall combines two mechanisms: semantic (vector) search, which finds statements whose meaning matches a query even when the wording differs, and graph traversal, which finds statements connected by explicit relationships even when the wording shares nothing. The consensus is that neither alone is sufficient — semantic search misses what only relationships reveal; traversal misses what only meaning-similarity reveals.
Recency- and validity-aware scoring. Retrieval ranks candidates by relevance and by temporal standing, so a statement updated last week outweighs a stale one, and the known failure mode — memory accurate at write time, quietly wrong at query time — is engineered against directly.
Autonomous graph construction. The systems build their graphs automatically: a language model reads incoming conversation and data, extracts entities and relationships, and writes them into the graph with no human confirmation step.
Research refinements. Recent papers separate associative proximity from causal structure (so agents can reason about why, not just retrieve what), address memory contamination through write isolation, and treat the graph as the space an agent plans and acts over rather than a passive lookup table ("agentic graph retrieval").
The converged shape is the seed's shape, arrived at independently:
memory_events append-only log.The field arriving here is confirmation of the seed's bet. It also does Loomworks' market education for free: prospects are being taught, by well-funded vendors, that agent memory should be a temporal graph.
3.1 Hybrid retrieval (feeds the memory-equals-recall reframe). Loomworks has the graph half of the converged retrieval pattern and no semantic half. There is no vector or meaning-similarity layer over Memory. "Recall" in practice means a Companion answering "what do we know that bears on this?" — and the field's settled answer is that this requires both mechanisms. Recommendation: when the recall reframe is commissioned and scoped, hybrid retrieval is the opening architecture question, not a later refinement. Alternative considered and set aside: graph-traversal-only recall. Set aside because the field's production experience is consistent that traversal alone misses semantically relevant, structurally unconnected material — the exact material a human forgot to relate.
3.2 As-of-a-date queries (feeds the audit and compliance fit). The research formulation stores facts with start and end validity times so a query can return what was true at a specific moment, never mixing stale and current. Loomworks can already reconstruct any single assertion at any past moment, but the graph as a whole has no "as of the twelfth of March" query — relationship objects carry no validity window and no whole-graph point-in-time read exists. For the strongest-fit market (financial services compliance, audit, wealth management), "show the engagement's knowledge state exactly as the board saw it on the decision date" is a demonstrable, sellable capability. The event log already contains everything needed; this is a query-layer addition, not a schema rework. Candidate for the build queue, unscheduled.
3.3 Supersession-aware retrieval scoring (a guarantee-in-code candidate). The field scores retrieval by recency. Loomworks records supersession more rigorously than any scanned system but does not yet use that record at retrieval time. The rule to adopt: a Companion answering from Memory must be structurally prevented from surfacing a superseded assertion as current — enforced in the retrieval code, not requested in a prompt. Small scope, high value, consistent with the guarantees-in-code principle. Candidate for the build queue, unscheduled.
Every scanned production system builds its graph autonomously: extraction straight from conversation into the graph, no human gate. Adopting this would delete the human-attributable confirming party that Loom Protocol conformance requires at commit, and with it the entire trust model — Operator authority, attestation, the audit-grade record. This is the same bright line the OpenClaw comparison drew for eyes-and-hands agent access: the divergence is principled, not incidental. Their model is agent working notes at graph rigor; the Loomworks model is the governed record from which the engagement is derived.
Precision on where the line sits: extraction-assisted contribution is fine and already shipped — the machine may propose, extract, and hold. Extraction-driven commit is the refusal. The machine surfaces and signals; the Operator approves.
Per the elaboration document (Claude Code inspection, 2026-07-27): the backward and sideways walks are live; the forward walk into Shapes and Renders is unbuilt; the rejection relationship is declared but has no producer; redirect leaves no graph edge. The field scan does not change any of that — it raises the confidence that the forward walk and the recall layer are the right next investments, because they are exactly where the field's value claims concentrate.
The category now has a name the market recognizes — graph memory for agents — and vendors teaching its value. None of the scanned systems have human-gated commit, attestation at a substrate chokepoint, non-suppressible marking of machine origin, or recorded considerations. The differentiator to state plainly: they offer autonomous graph memory — fast, ungoverned, built by the machine for the machine; Loomworks offers governed graph memory — the same temporal-graph substrate, under Operator authority, with every commit attributable to a person and every transition attestable. Where the work matters — compliance, audit, wealth management, legal, healthcare — governed is the version that survives scrutiny.
Field claims draw on material current to 2026-07-27: the Graphiti open-source repository (getzep/graphiti) and Zep's temporal-knowledge-graph architecture paper (arXiv 2501.13956); Supermemory's temporal-knowledge-graphs-for-agent-memory overview; Mem0's graph-memory-solutions comparison (covering Zep, Letta, Mem0); Neo4j's Graphiti write-up; and 2026 research preprints — MAGMA (arXiv 2604.12285 citations), GAM, AriGraph, temporal semantic memory (arXiv 2601.07468), and the agentic graph-retrieval survey line (via arXiv 2605.27864). Engine-state claims draw on loomworks-graph-role-elaboration-v0_1.md (loomworks-record, commit 3321fa6).
DUNIN7 — Done In Seven LLC — Miami, Florida Loomworks — Graph-Memory Field Scan — v0.1 — 2026-07-27