Coding agents like Claude Code, OpenAI Codex, and Google Antigravity are suffering from a specific kind of amnesia that costs developers hours of wasted compute. Alexander Panasenko, author of the open-source project KYB (Know Your Business), detailed in a recent blog post how scattered Markdown notes caused his agents to repeatedly pursue architectural paths that had already been rejected. The solution, pinned to KYB v0.3.3 released on October 27, 2026, is a Git-backed shared store that enforces versioned records with explicit lifecycle states, ensuring that 'active' knowledge is distinct from 'archived' work or 'retracted' errors.

The Stale Note Trap

Before implementing KYB, Panasenko relied on Markdown files scattered across workspaces. The critical failure mode occurred when a new agent session would encounter an older note describing a valid but abandoned strategy. Because the file lacked metadata indicating its obsolescence, the agent would spend hours re-implementing a design the team had already moved past. Panasenko notes that while Git tracks file changes, an agent reading a working tree cannot easily distinguish between a superseded decision and the current authoritative state without separate, manual checks. The 'append-only' habit of adding clarifications to the end of documents also failed, as agents often matched the first occurrence of a keyword rather than the most recent correction.

Lifecycle States and Temporal Slices

KYB v0.3.3 introduces a rigid distinction between three record types: Knowledge (state/rules), Incidents (problems/outcomes), and Tasks (work/commitments). Each record exists as a Markdown file in a Git repository, but the system adds query modes to handle time. A standard query returns the current state, while --as-of builds a slice of the store at a specific commit, and --history reveals the chain of changes. This prevents the 'write time vs. event time' confusion; for example, if a service port changed at 15:50 but was recorded at 16:20, a query for 16:00 correctly returns the old port, reflecting what the store knew at that moment. Retractions are explicit: a retracted record disappears from active queries but remains in Git history, signaling to the agent that the knowledge was wrong, not just finished.

Agent Feedback and Integration Gaps

Despite the robust backend, a qualitative survey of five internal coding agents on October 10, 2026, revealed a persistent integration gap: agents often forgot to call the memory tool. While KYB provides CLI, HTTP, and MCP interfaces, some memory files still loaded automatically into context, whereas KYB required an explicit query. Agents requested a 'short digest' of workspace records at session start and explicit verification status for facts, such as who confirmed a configuration and at which revision. The v0.3.3 release added deadline tracking with due and hard fields, but it does not run tasks on a schedule or send proactive notifications; the agent must still actively retrieve the reminder. This highlights a fundamental truth in agent architecture: having a path to correct knowledge is useless if the agent lacks the prompt or habit to take it.

Key Takeaways

  • KYB v0.3.3 uses Git commit history to provide temporal query modes (--as-of, --history) that distinguish current truth from past attempts.
  • The system enforces lifecycle states (active, archived, retracted) to prevent agents from mistaking finished work for current knowledge or wrong answers for valid data.
  • A survey of agents revealed that explicit tool calls are often skipped in favor of automatic context loading, leading to partial knowledge retention.
  • Deadline features in v0.3.3 provide visibility but do not automate task execution or notifications, requiring active agent engagement.

The Bottom Line

Memory isn't just storage; it's about state management. Until agents have automatic context injection that doesn't bloat tokens, tools like KYB are essential guardrails against the expensive cycle of re-learning rejected decisions.