Cognition OS runs housekeeping agents that keep it coherent: merge / split for nodes that are the same thing or conflate two, link prediction for edges the data implies but nobody stated, hierarchy coherence for parentage that drifted, concept drift for regions that stopped matching their own summary, and quality for naming and structure. Each runs under its own timeout, and because housekeeping mutates the graph, every change is recorded and attributable.
Provenance is the other half. Cognition OS keeps a Graph of Thought: observations, inferences, decisions, questions and answers as a directed acyclic graph, so you can ask any entity what produced it, and creations, merges, splits and links all point back to their cause. Reasoning records the same chain on its side.
On the memory side, Kemory keeps capture separate from consolidation: capture stores discrete items the user confirmed, while consolidate only refreshes a session summary and stores nothing new. Its results say how a search actually ran, and the documentation tells you to read them honestly: an empty result in encrypted mode is not evidence that a memory does not exist. The Kemory panel for Claude Code puts that in front of a person: which memories are in context, with a place to review, rate and save them.
Forgetting is part of the design. Kemory’s memories age through normal decay to an archive, and The Librarian in Your Head (October 2025) sets out the design patterns behind that idea: compress a long history into a summary that keeps its citations, and down-weight a memory that keeps misleading, flagging it for review.
The most transferable habit is publishing what failed. Kemory’s retrieval page lists fixes that were built, measured against a labelled set, and rejected: chunking before embedding (actively harmful), cross-encoder re-rankers (no measurable effect), an asymmetric query prefix (slightly worse) and a fixed similarity floor (inert at scale). The one thing that consistently helped was phrasing queries in the vocabulary the memory was written in.
From SeKondBrain research
- Put time and certainty in the data model. Concept-Grounded Attention (September 2026) put the concept graph inside a transformer and tested each mechanism against itself switched off. Steering attention with graph structure changed nothing, and concept-graph retrieval recovered multi-hop paths without improving evidence recall. What helped was the memory model underneath: 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%. For a graph builder: give facts valid-from dates and a confidence or status before reaching for cleverer retrieval.
- Measure consolidation, and say where it falls short. Atomic Units of X frames intelligence as organising evidence into reusable atoms. Its tests on three open-source software projects hit the target ratio on two and missed on the third, and the paper reports the miss as plainly as the wins. The research page summarises both.
- Every claim should beat its own off-switch. Both papers run the same discipline the retrieval notes do: implement, measure against a baseline, publish the negative result.
- Edges for meaning, distance for strength (research direction). Context Graphs Are the Bridge describes a direction SeKondBrain is exploring: typed edges say what a relationship is, and distance says how strong or relevant it is. Kemory already uses the first half in production: proximity in the memory graph is one of its five re-ranking signals. The geometry half is research, not a Cognition OS or Kemory feature.
Going further: hypergraphs and hyperbolic graphs
Optional, and early. 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 benchmarks, not yet independently replicated; the event node in step 1 gets you most of the way today. Hyperbolic embeddings place nodes in curved space where deep hierarchies, such as a label chain, 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 research, not a product feature, and you do not need them to build a graph.