Knowl · persistent memory for AI coding agents

Stale knowledge is worse than none.

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 — the approaches that failed first, and the measurements that ruled each one out.

~/your-repo

$ npm install -g @dat999zx/knowl
added 1 package
$ knowl --version
5.23.1

Node 20 or newer. No API key, no vector database, no service to run alongside it.

Forty seconds on what Knowl does.

Knowledge doesn’t get lost. It goes quietly out of date.

In March an agent works out your API rate limit. 100 a minute.

In June somebody raises it to 500. The note still says 100.

An agent reads the note. It builds against 100.

Nothing announces it. You find out when something breaks.

Memory that survives the session

Four facts in a chain, two checked, one crossed and pointed at

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

Facts held locally, one of them published upward to a cloud

Yours until you publish

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

Two faded facts replaced in turn by a current one

Built to go out of date

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 Superseded knowledge is never deleted — there is no delete verb in the product at all.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

Get started

Adapters

Five clients. One command each.

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.

Claude Codeknowl init claudeMCP + hooks
Codex CLIknowl init codexMCP + hooks
Cursorknowl init cursorMCP + hooks
GitHub Copilotknowl init copilotMCP + hooks
OpenHandsknowl init openhandsMCP + hooks
Antigravityknowl init antigravityMCP + hooks
Windsurfknowl init windsurfMCP + gate
Clineknowl init clineMCP + plugin
Zedknowl acpMCP + ACP
JetBrainsknowl acpMCP + ACP
Neovimknowl acpMCP + ACP
Kiroknowl acpMCP + ACP
Claude Desktopknowl init claude-desktopMCP
OpenCodeknowl init opencodeMCP
Anything elsespeaks MCPbring your own

Read the adapter docs

Before / after

What it replaces

Most teams already have agent memory. It is a markdown file nobody trusts, and a person who remembers the rest.

CLAUDE.mdAPI rate limit: 100/minSomewhere in nine hundred lines. True in March, still there in June.
The fact, and its successor500/min supersedes the March factThe old one stays readable. Nobody had to remember to edit anything.
Pasting context every sessionThe same explanation, every new chat
The agent already knowsAnd can show you where it learned it
Asking in Slack“does anyone remember why we…”
One workspace memoryDisagreements surface as disputes, not silence
A teammate who remembersUntil the day they don't, or they leave
A record that outlives themWith the alternatives it beat still attached

Start free

Freshness

One fact, from written down to out of date

A memory tool that only stores things goes stale silently. That rate limit, in full: what Knowl records over the four months it takes one claim to go from true to retired — every step kept, nothing overwritten, nothing deleted.

In MarchYour agent asks
“What's our rate limit on the public API?”
100 requests a minutefactobservedapi-gateway
An agent worked this out mid-task and stored it, with the run that established it still attached.
A week laterA teammate confirms it
Approved in the review queue
From here the claim carries a name, not just a timestamp — so “who decided this” still has an answer six months later.
In JuneThe same question
“What's our rate limit on the public API?”
500 requests a minutesupersedes the March fact
It changed, and the answer changed with it. Nobody had to remember to go and edit anything.
Any timeAsk what it used to be
The March answer is still there
Retired, not deleted. You can see that it changed, when, and what replaced it — which is the part a file full of notes can never tell you.

Proof

The right answer, even after it changed

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.

90%Knowl

79%agentmemory

60%GPT-4olong-context

54%HippoRAG-v2

48%BM25

45%GPT-4o-minilong-context

29%Qwen3Embedding-4B

28%Cognee

28%MemGPT

18%Mem0

14%MIRIX

7%Zep

SubEM on 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.

In their words

The people shipping agent memory say the same thing

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

OpenAI

“Memory can become stale. Agents are instructed to treat memories as guidance only.”

Agents SDK — agent memory

Greyhound Research

“The deepest flaw in current agent memory is that it mistakes retrieval for memory.”

Sanchit Vir Gogia, Chief Analyst — via InfoWorld

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

You pay to publish, never to read

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.