--- title: "What is a company brain? Timeline and requirements" description: "A definition of the company brain, its main components, and the memory functions required to keep company information current." canonical: https://past.dev/blog/company-brain date: 2026-08-30 category: Research authors: The past.dev team --- # What is a company brain? Timeline and requirements A company brain collects information from documents, email, chat, tickets, meetings and business systems. It makes that information available to employees and AI agents through search, chat or an API. Y Combinator listed the Company Brain in its [Summer 2026 Requests for Startups](https://www.ycombinator.com/rfs?year=2026). Related work uses terms such as GBrain, context graph, LLM wiki and agent memory. Each term covers a different part of the system. ## Company brain timeline The phrase “company brain” is older than the current AI product category. From 2022 to 2026, work in personal knowledge management, agent memory, graph retrieval and enterprise decision records defined its current meaning. - **June 2022.** Tiago Forte published [Building a Second Brain](https://www.buildingasecondbrain.com/). It describes a personal system for collecting, organizing and reusing notes and reference material. - **2023.** The [Generative Agents](https://arxiv.org/abs/2304.03442) paper gave agents a memory stream with retrieval, reflection and planning. [MemGPT](https://arxiv.org/abs/2310.08560) described memory tiers that move information between an LLM's context window and external storage. - **April 2024.** Microsoft researchers published [GraphRAG](https://www.microsoft.com/en-us/research/publication/from-local-to-global-a-graph-rag-approach-to-query-focused-summarization/), a graph-based method for answering broad questions over large private text collections. - **December 22, 2025.** Jaya Gupta and Ashu Garg published the [context graph](https://foundationcapital.com/ideas/context-graphs-ais-trillion-dollar-opportunity) thesis. It describes decision traces that connect rules, exceptions, approvals and precedent across systems, entities and time. - **April 4, 2026.** Andrej Karpathy published the [LLM Wiki](https://gist.github.com/karpathy/442a6bf555914893e9891c11519de94f) pattern for personal knowledge bases. An agent converts source material into linked Markdown pages and updates those pages as new information arrives. - **April 2026.** Garry Tan released [GBrain](https://github.com/garrytan/gbrain), the open-source knowledge system used by his AI agents. It ingests records, creates pages for people and companies, and supports search, citations and synthesis. - **Summer 2026.** Tom Blomfield made the Company Brain a named category in [Y Combinator's Requests for Startups](https://www.ycombinator.com/rfs?year=2026). YC described a system that collects fragmented company knowledge, structures it, keeps it current and makes it usable by people and AI agents. Together, these projects define the current category: ingest company records, resolve the people and organizations they mention, track changes, return supporting sources and provide the same information to employees and agents. ## Company brain requirements A company brain should meet four requirements. - **Current information.** Answers should use the latest supported facts. - **Sources.** Answers should include the source and date for each fact. - **Coverage.** The system should ingest the tools where work occurs, including email, chat, tickets and meetings. - **Programmatic access.** AI agents should be able to query the same information through an API. Keeping information current requires an explicit data model for changes. Manual documentation updates are not sufficient for high-volume activity data. ## Components A company brain usually contains five processing stages. 1. **Ingestion.** Import activity from company systems. Preserve the original timestamp for every record. 2. **Extraction.** Identify facts, events and relationships. Retain a link to the source. 3. **Entity resolution.** Match references to the same person, company or project across systems. 4. **Temporal storage.** Record when each fact was valid. Retain previous values and conflicts. 5. **Recall and interfaces.** Provide evidence through an API for agents and through search or chat for employees. The [agentic memory article](/blog/agentic-memory) describes these processing stages in more detail. ## Current-state accuracy Document retrieval can return an outdated statement when an older page closely matches a question. Current-state answers require additional data. - **Resolved entities** connect references across email, transcripts and chat. - **Validity dates** record when a fact became current and when it was superseded. - **Conflict records** preserve disagreements between credible sources. - **Insufficient-evidence status** lets the application report that the records do not support an answer. These functions allow the system to answer current and historical questions from the same records. ## Implementation options Teams can use company-brain features in an existing software suite, buy a dedicated product, build an internal system or add memory infrastructure to their own application. The [agent memory comparison](/blog/mem0-alternatives) covers several memory products. past.dev provides memory infrastructure through an API. `POST /api/v1/ingest` accepts raw text and its original timestamp. past.dev performs extraction and entity resolution. `POST /api/v1/recall` returns ranked, dated evidence with a `status` value such as `Supported`, `Conflicted` or `NoKnownSupport`. The [benchmarks page](/benchmarks) documents the method and results. The [quickstart](/docs/memory-api/quickstart) documents the API calls. [Context engineering](/context-engineering) explains how an agent can use the returned evidence.