Show HN: Knowl – agent memory that retires stale facts when they change
Knowl is an open-source agent memory library that flags outdated facts as "superseded" at write-time rather than accumulating conflicting entries.
Knowl, an open-source agent memory tool built to automatically retire stale facts when they are overwritten, launched on Hacker News on August 22, 2026 under a Show HN post by author dat999zx, per the HN submission.
The project addresses a specific failure mode in long-running AI agent sessions: existing memory systems append new information without resolving conflicts with older entries, leaving agents unable to determine which fact is current. The GitHub repository describes how Knowl splits knowledge into discrete units called "atoms," each typed as one of: fact, decision, goal, constraint, architecture, state, or skill. When a new atom conflicts with an existing one at write-time, the old atom is flagged as superseded and removed from the main retrieval index — though the full history is preserved. The creator illustrates the problem with a concrete example: telling the agent on day one to use Lemon Squeezy as a merchant of record, then switching to Polar on day two — a scenario where append-only memory systems return both answers or fail to pick one.
The tool connects to AI coding environments through MCP (Model Context Protocol) and is listed as compatible with Claude Code, Codex, Cursor, and Antigravity. It runs fully locally by default. A hosted option, Knowl Cloud, is offered for team synchronization. Additional features noted in the HN post include transcript search, multi-workspace sharing, and change-detection impact analysis.
The creator benchmarked Knowl on MemoryAgentBench FactConsolidation single-hop at 262K context, reporting a score of 0.90, versus 0.79 for agentmemory, 0.60 for GPT-4o on full context, 0.18 for Mem0, and 0.07 for Zep. On the multi-hop task, Knowl scored 0.07, which the author notes matches the benchmark's reported ceiling of 0.14. The benchmarks were run at temperature 0.7, and full results are available in the repository. These figures are self-reported by the author and have not been independently verified.
No company name, funding, team size, or founding date is disclosed in the source. The creator is soliciting community feedback and positions the project as fully open-source, with Knowl Cloud as an optional add-on for teams.
No comments yet — start the thread.