Show HN: Knowl – agent memory with write-time supersession, 0.90 on MAB

Knowl, an agent memory tool that runs as an MCP server, announced it scored 90 out of 100 on the MemoryAgentBench FactConsolidation-SH benchmark at 262K, addressing the problem of stale agent memory. The tool, version 5.14.0, records reasoning, evidence, and provenance for every fact, and supersedes outdated information without deleting history, with integrations for Claude, Codex, Cursor, Copilot, and others.

Checkable by design Every fact carries its reasoning, its evidence and its provenance — observed, stated by a person, or inferred. A claim you can check without asking anybody. Reasoning, evidence and provenance on every fact Because it gets trusted. Knowl gives every fact your agents store a record of where it came from , a review trail , and a successor when it changes — so the answer they hand back is the current one, and you can see who settled it. How this was built https://blog.knowl.cloud/the-story-of-knowl — the approaches that failed first, and the measurements that ruled each one out. ~/your-repo bash $ npm install -g @dat999zx/knowl added 1 package $ knowl --version 5.14.0 Node 20 or newer. Nothing else to install. A session ends and the context goes with it. It gets re-proposed. Six months on, nobody can say why it was decided. Somebody acts on it. Stale knowledge is worse than none, because it gets trusted. The next one starts from zero. The same ground gets relearned every week. Every fact carries its reasoning, its evidence and its provenance — observed, stated by a person, or inferred. A claim you can check without asking anybody. Reasoning, evidence and provenance on every fact Knowl runs in your repository and keeps its knowledge there. Your code, your embeddings and your session history are never uploaded. Local by default, shared on purpose A fact records what it replaced and what replaced it. Superseded knowledge is never deleted — there is no delete verb in the product at all. History is the feature, not a changelog Adapters Knowl runs as an MCP server, so anything that speaks MCP works with no adapter to write. Three of these also take lifecycle hooks, which is what lets an agent capture what it learned without being asked. knowl init claude MCP + hooks knowl init codex MCP + hooks knowl init cursor MCP + hooks knowl init copilot MCP + hooks knowl init openhands MCP + hooks knowl init antigravity MCP + hooks knowl init windsurf MCP + gate knowl init cline MCP + plugin knowl acp MCP + ACP knowl acp MCP + ACP knowl acp MCP + ACP knowl acp MCP + ACP knowl init claude-desktop MCP knowl init opencode MCP speaks MCP bring your ownBefore / after Most teams already have agent memory. It is a markdown file nobody trusts, and a person who remembers the rest. Freshness A memory tool that only stores things goes stale silently. This is what happens to a single claim over four months — every step recorded, nothing overwritten, nothing deleted. “What's our rate limit on the public API?” 100 requests a minute fact observed api-gateway Approved in the review queue “What's our rate limit on the public API?” 500 requests a minute supersedes the March fact The March answer is still there Proof When a fact changes, most memory tools keep handing back the old one. On the public benchmark for exactly this, Knowl got 90 of 100 right. SubEM on MemoryAgentBench https://github.com/HUST-AI-HYZ/MemoryAgentBench FactConsolidation-SH at 262K, scored with the benchmark’s own harness. Method, raw results, and the splits where we do worse: the full write-up https://github.com/dat999zx/knowl/blob/main/benchmarks/memoryagentbench/mab/FINDINGS.md . In their words Memory going stale is not our framing of the problem. It is written into the documentation of the agents your team already runs. Anthropic “Over many sessions a memory store accumulates duplicates, contradictions, and stale entries.” Managed Agents documentation — Dreams https://platform.claude.com/docs/en/managed-agents/dreams OpenAI “Memory can become stale. Agents are instructed to treat memories as guidance only.” Agents SDK — agent memory https://openai.github.io/openai-agents-python/sandbox/memory/ Greyhound Research “The deepest flaw in current agent memory is that it mistakes retrieval for memory.” Sanchit Vir Gogia, Chief Analyst — via InfoWorld https://www.infoworld.com/article/4191031/microsoft-unveils-memora-to-tackle-ai-agents-memory-problem.html Quoted verbatim from published sources, read August 2026. These are observations about the problem, not endorsements of Knowl — every link goes to the page the sentence was read on. Pricing Queries never reach us — your agents read a replica on your own machine, so reading back is free at every tier and always will be. Accounts are free and so is your first workspace. Anyone can be invited in and use all of it — you pay when your team outgrows what Free carries each period. A hosted memory layer for coding agents. Your agents write what they work out to a local store; publishing lifts a fact into a shared workspace your team reads from. No. The index, the embeddings and every session stay on your machine. Publishing is an explicit act, and only the fact you publish travels. It is superseded, not deleted. The new fact records what it replaced, the old one stays queryable as history, and only the current answer is returned by default. A wiki has no idea when it went stale, and a vector store returns the old claim beside the new one. Knowl records supersession, so the answer you get is the current one. Claude Code, Cursor, Codex, Gemini CLI and Claude Desktop, configured by one command. Anything that speaks MCP works without an adapter. Only the members of the workspace you published into. Knowledge written privately stays on your machine until someone promotes it deliberately. Every write is secret-scanned on the server, on the way in, every time — not once at setup. A refused write names what tripped it without echoing the value back. One repository is enough. Connect it, sign the CLI in, and let a week of work accumulate before you judge it.