3D agent memory critique
An engineer laid out a '3D' view of agent memory—relational, vector and graph stores—and argued single-vector DBs are limited for business knowledge like outage timelines; they promoted an OSS called Cognee to unify stores (x.com/i/status/2044329897603244093). The thread emphasized mixing storage types to represent temporal events, relations and embeddings together for more reliable agent memory retrieval (x.com/i/status/2044329897603244093).
A debate over how artificial intelligence agents should remember information is shifting from “use a vector database” to “use several kinds of storage at once.” (docs.cognee.ai) A vector database stores embeddings, or numerical fingerprints of meaning, so a system can find text that is conceptually similar even when the wording changes. Cognee’s documentation says that works for semantic search, while a relational store tracks source documents and provenance and a graph store maps entities and relationships. (docs.cognee.ai) Cognee described that split in its architecture docs and in its public codebase, which showed about 15,600 GitHub stars and roughly 1,600 forks on April 17, 2026. Its docs say the platform can be configured with separate embedding engines, vector stores, graph databases and relational backends. (github.com, docs.cognee.ai) The critique behind the recent thread is that a single semantic index is good at “find me something similar,” but weaker at “what happened first, what caused it, and what system was involved.” Cognee’s architecture page says no single database handles all aspects of memory and that hybrid search combines semantic and structural retrieval. (docs.cognee.ai) That distinction matters for business records that behave more like timelines than isolated paragraphs. A service outage, for example, can include a start time, a repair step, a dependency between systems and a postmortem source, which fit naturally into tables and graphs as well as embeddings. (docs.cognee.ai) The broader agent market is also moving toward explicit state management rather than treating memory as a single search box. OpenAI’s Agents SDK documentation says agents “keep enough state to complete multi-step work” and highlights custom storage and server-managed conversation strategies for developers who own orchestration. (developers.openai.com) Earlier retrieval patterns often centered on chunking documents, generating embeddings and querying a vector backend. OpenAI’s Retrieval Plugin repository describes that design as a retrieval system for personal or organizational documents, with document embeddings stored and queried through a vector database. (github.com) Cognee’s pitch is not that vectors are unnecessary. Its docs say the vector store still handles semantic similarity, while the graph store supports structural searches through entities and links and the relational layer preserves provenance about where each chunk came from. (docs.cognee.ai) The company’s public documentation also says the same information may be indexed in more than one store for efficiency. That means one memory item can be searchable by meaning, tied to a source record and linked to related entities at the same time. (docs.cognee.ai) The thread’s core claim is less about replacing one database with another than about matching storage to the shape of the question. If an agent needs to recall not just similar text but sequence, source and relationships, the memory system has to store those facts in forms built for each job. (docs.cognee.ai, developers.openai.com)
