{
  "id": 5165136,
  "title": "Almanac's Company-Context Agent: How YC S26 Wires Internal Knowledge into Every LLM Call",
  "url": "https://urgent.news/2026/09/02/almanacs-company-context-agent-how-yc-s26-wires-internal-knowledge",
  "topic": "ai",
  "section": "AI",
  "published": "2026-09-02T20:07:20.000Z",
  "source": {
    "name": "Dev.to",
    "slug": "dev-to",
    "url": "https://dev.to/mech_app_ai/almanacs-company-context-agent-how-yc-s26-wires-internal-knowledge-into-every-llm-call-45gc"
  },
  "original_language": "en",
  "account": "Almanac, a startup from Y Combinator S26, has developed a unique solution to the context-persistence challenge faced by multi-agent financial systems. Their product aims to eliminate the problem of agents forgetting company policies between interactions, hallucinating organization structures, or repeating questions.\n\nAt the core of Almanac's approach is a live wiki that compiles company knowledge from connected tools such as Slack, email, and project management platforms. This wiki serves as the source of truth for the agent, ensuring that the information is always up-to-date and relevant.\n\nThe retrieval pipeline architecture consists of three layers: tool connectors, wiki compiler, and context selector. Tool connectors use OAuth integrations to stream events from various sources and tag each event with metadata such as author, timestamp, project, and customer name. The wiki compiler groups related events into pages, with pages representing individual customers, projects, or other relevant entities. The wiki compiler runs continuously to ensure the wiki remains current.\n\nWhen the agent receives a task, the context selector queries the wiki using a combination of keyword matching and semantic search. It returns a ranked list of relevant pages for the agent to read and review. If the agent requires additional context, it can request further information, triggering a new retrieval pass with updated token budget calculations.\n\nPermission boundaries are enforced through two main points: at ingestion and at retrieval. During ingestion, the wiki compiler inherits the access control list (ACL) from the source tool, ensuring that only authorized information is included in the wiki. At retrieval, the context selector filters pages based on the user's permissions, preventing sensitive data from being exposed to unauthorized users.\n\nAlmanac's token budget strategy is designed to balance company context with user queries and agent reasoning. The system allocates a reserved buffer of 1,000 tokens for overflow situations, such as long tool outputs or multi-turn conversations. If the top three wiki pages exceed the allocated tokens, the system summarizes the pages using a separate LLM call before passing them to the agent.\n\nTo address context drift and maintain accuracy, Almanac handles changes to company policies through event-driven updates. When a Slack message is edited, an email is recalled, or a GitHub issue is closed, the wiki compiler re-processes the affected thread and updates the relevant page in real time. This ensures that the agent's knowledge remains current and relevant, minimizing the impact of context drift on future interactions.",
  "summary": "Almanac (YC S26) launched with a promise to solve the context-persistence problem that breaks most multi-agent financial systems: agents that forget company policies between calls, hallucinate org structure, or ask the same questions twice. Their pitch is \"Hermes with a brain,\" an agent that maintains a self-updating wiki of company knowledge and injects relevant context into every LLM call. The…",
  "key_points": [],
  "editors_take": null,
  "illustration": null,
  "coverage": {
    "outlets": 1,
    "also_reported_by": []
  },
  "ai_generated": true,
  "disclaimer": "Summaries, key points and the editor’s take are written by software from other outlets’ reporting and may contain errors — always check the linked original."
}