Slack Adopts Structured Memory for Multi-agent Context

InfoQ reports that Slack engineers shifted from accumulating raw chat logs toward using structured memory, validation, and credibility-weighted findings to manage coherence in long-running multi-agent systems. InfoQ quotes the observation that "Agent frameworks solve the state management problem for users by accumulating message history between API calls," and warns that approaching an agent's context window limit degrades response quality. InfoQ reports Slack staff software engineer Dominic Marks described one application that spans hundreds of requests and generates megabytes of output. InfoQ describes Slack's three context channels as a director's journal, a critic's review, and a critic's timeline, and reports that a central coordinator dispatches work to specialized experts and critics.

InfoQ reports the three channels serve distinct roles: the director's journal stores the director's structured working memory including findings, observations, decisions, questions, and hypotheses; the critic's review stores an annotated findings report with credibility scores derived from evidence inspection; and the critic's timeline stores chronological findings with credibility annotations. InfoQ describes critics as receiving summaries from experts, evaluating evidence to build a credibility-weighted list of corroborated findings, and producing scores used to filter or prioritize outputs.

Industry-pattern observations: separating persistent, structured memory from transient chat history is a common approach to scale agent state without repeatedly filling context windows. Systems that attach provenance and credibility scores to evidence help triage outputs and reduce the impact of hallucinations, while a coordinator-and-specialist topology simplifies role separation between generation and verification.

for practitioners building long-running agentic workflows, Slack's pattern emphasizes three engineering trade-offs: explicit memory schema design, automated validation/credibility estimation, and orchestration overhead for coordinator/critic interactions. These patterns address scaling and reliability challenges that arise once sessions stretch beyond single-request interactions.

observers should watch for open-source toolkits, standard formats for structured agent memory, evaluation metrics for credibility scoring, and how latency and cost trade-offs evolve when adding critic layers.