A graph that only grows becomes a landfill. @0xWast3’s Memory Engineering (4 Aug 2026) gives the maintenance loop:
- Capture only facts that would still matter in three months.
- Consolidate duplicates into one node.
- Retrieve the top few relevant nodes for a question, not the whole graph.
- Reconcile disagreements: supersede (newer wins, archive the old), coexist (both true in different contexts), or flag the conflict.
- Decay unused nodes by archiving them, never deleting. When you compress an old history into a summary node, keep links to the facts it summarises so the summary still has receipts, and down-weight a memory that keeps misleading (The Librarian in Your Head).
Deletion is a chore too. When a source is withdrawn, every fact built from it should go with it. Lightfoot (arXiv listing 10 Nov 2026) shows why provenance makes that easy: because every entity, link, and fact is built from per-passage extractions that remember their chunk, deleting a document is a database operation, and the graph afterwards matches a fresh rebuild from the surviving documents.
Apply those rules to the memory store you already have; do not build a custom vector memory engine. Graphiti and iText2KG do the reconcile step with dates on edges (a new fact can sit beside an older observation). The RAG playbook and CareerStack add the weekly chores: look for orphan nodes and near-duplicates, version the schema, and rerun the competency questions after every change.
From recent research
The 15 and 16 Nov 2026 updates draw on recent agent-memory and graph-construction papers that earn a place here because they teach a construction or maintenance habit, not because they were on a news wire.
Keep the receipts. Can LLM Agents Organize Their Own Memory? (Fu et al., arXiv listing 10 Nov 2026) has the agent summarize, group related information, and revise topics, with every summary still linked to the original interactions as evidence. Discovering Effective Agent Memory Designs through Program Evolution (Ding et al., arXiv listing 13 Nov 2026) finds the same pattern from the other direction: keep raw conversations as retrievable evidence, because extracted memories drop details, and compose that evidence according to the query's intent. Don't Just Trust Your Memory (Lee et al., arXiv listing 11 Nov 2026) shows why the link matters: a stored lesson can be wrong, or right but misapplied, and the agent then fails a task it would have solved without memory.
Write first, resolve later. DISCERN (Ahuja et al., arXiv listing 13 Nov 2026) argues that persistent memory is not mainly a retrieval problem: preserve what you write, and resolve conflicts when you read. That is the same spirit as Graphiti's dated edges.
Do not treat a stored memory as automatically true. MemoryDojo (Yang et al., arXiv listing 9 Nov 2026) finds that memory systems can be poisoned, with substantial variation in the damage. In a governed setting, Runtime Auditability for Healthcare RAG Systems (Chockalingam, SSRN posted 13 Nov 2026) makes every decision to admit, assemble, or release content a logged decision, with a source trace and a human attestation a third party can examine.
Build the graph so it can explain and undo itself. AutoOntic (arXiv listing 13 Nov 2026) shows what open extraction leaves behind, a flat graph with no hierarchy and no rules about which node may join which, and recovers a model from it. TripleGrader (arXiv listing 10 Nov 2026) puts a number on trusting raw extraction: most LLM triples needed a fix. The Vietnamese Legal Ontology (arXiv listing 10 Nov 2026) keeps the source record and method on every node and edge, and Lightfoot (arXiv listing 10 Nov 2026) uses the same habit to delete a document cleanly.
Put time and certainty in the data. Concept-Grounded Attention (Duggal, Swarnalatha Ramanna and Vassiliades, arXiv, September 2026) tested a concept graph inside a transformer against its own switched-off controls. Steering attention with graph structure added little; explicit time added 13 to 25 points on LongMemEval across generators up to 122B parameters, and an epistemic status on each claim cut unsupported assertions in 72–122B models from 19–68% to 0–5%.
Extract with a scope, and retrieve with more than one method. Who Remembers What? (Wang et al., arXiv listing 10 Nov 2026) is about multi-user, multi-agent memory: keep stated preferences, but keep private facts inside the scope they belong to. When does Hybrid Retrieval Help? (Tallam et al., SSRN posted 13 Nov 2026) surveys combining lexical and semantic retrievers rather than betting on one. And don't pour the whole graph into the prompt: CompactionBench (Liu et al., Microsoft Research, 12 Nov 2026) is a reminder that long-horizon agents which cannot compact their context waste tokens and time; retrieve the few relevant nodes, or compact what must stay in context.
Going further: hypergraphs and hyperbolic graphs
Optional, and early. Native hypergraph retrieval stores many-part facts as hyperedges: HyperGraphRAG (Luo et al., 2025) and Hyper-RAG (Feng et al., 2025) report gains over standard RAG and GraphRAG on their own domain benchmarks, not yet independently replicated; the event node above gets you most of the way in any store today. Hyperbolic embeddings place nodes in curved space where deep hierarchies fit with little distortion (Poincaré embeddings, Nickel & Kiela 2017; MuRP, 2019; AttH, Chami et al. 2020). They help link prediction on hierarchy-heavy graphs, mostly at low dimensions. They are a research and embedding-layer technique, not a database feature, and you do not need them to build a graph.