Company memory for AI agents

your agents forget everything. witz remembers.

Connect your company's tools once. Choose exactly what your agents may know. Every agent — Claude Code, Cursor, Codex — reads the same trusted, cited context from one MCP endpoint, and you can see every read.

Works with · Claude Code · Cursor · Codex · Claude · ChatGPT
Claude Code · asked witz
“Why did we switch webhook retries to exponential backoff?”
A June incident showed fixed 30s retries stampeding a recovering endpoint. The team moved to exponential backoff with jitter, capped at 5 attempts.
↳ decision · webhook retry ↳ incident · PAY-214
memory_search · 2 fragments · 7ms · logged
1MCP endpoint every agent reads
6MCP tools, one connection
100%of claims carry a citation
<60sto your first cited answer
The problem

your agents have amnesia.

Every agent you run knows the code in front of it — and nothing about the specs in Notion, the tickets in Linear, or the decisions buried in old pull requests. So it guesses, or asks you to re-explain the company every session.

01
Wiring each tool separately floods the agent with hundreds of raw API schemas it can't reason over.
02
Nothing links or ages. A stale price and a current one look identical; contradictions get served as fact.
03
No scope, no record. You can't say “agents may read this, not that,” and you can't see what they read.
The mental model

Think of it as a second brain for your company — except the reader is your AI agent, not you.

You curate what goes in and keep it clean. Witz turns your tools into cited fragments, links them into a graph, and serves the right Memory to each agent over one MCP endpoint. Every read lands in a log you can audit.

three steps to memory.

Connect · curate · serve
01 · connect & select

You choose what's in

Link Linear, Notion or GitHub read-only. A selection tree lets you tick exactly which projects, spaces and repos become memory — nothing you didn't pick, sensitive containers out by default.

Screen · Sources
02 · witz builds memory

Clean, linked, dated

Your documents become concise fragments, linked into a graph of people, projects and decisions. Facts carry validity windows, so stale and conflicting answers are caught, not served.

Screen · Graph · Fragments
03 · agents read it

One endpoint, cited

Compose scoped Memories and paste one config into any agent. Answers come back cited to your documents — and every read lands in a log you can audit.

Screen · Memories · Connect · Retrieval log
A day with witz

decision archaeology.

A new engineer picks up a payments ticket and asks their agent why the code looks the way it does. Without Witz, the agent shrugs. With it, the answer is one query away — sourced, dated, done.

terminal · claude code
# new hire, first payments ticket
why does the decline-retry path reuse the original charge?
A declined-then-retried card once created a second charge. The retry path now reuses the original charge with an idempotency key, so a retry can never double-bill.
↳ issue · PAY-203 ↳ decision · idempotent retries
memory_search · memory: coding · 2 fragments · 8ms · logged ✓

what you get.

Six promises
01

Cited answers

Every claim links back to the document it came from, with a date. A fragment without evidence is shown as an audit item, never as truth.

02

One endpoint

Six MCP tools, one connection. Skip the per-tool sprawl and the raw-JSON flood in your agent's context.

03

You choose what's shared

Selection comes before ingestion, and scoped Memories per agent. Nothing leaks by default; source code is never indexed.

04

Every read logged

The retrieval log shows what each agent read, when, via which grant, and how fast.

05

Conflicts surface

Facts carry validity windows and version chains. When sources disagree, Witz shows the conflict instead of averaging it.

06

Works with every agent

Claude Code, Cursor, Codex, Claude and ChatGPT all read the same memory. No lock-in.

Run it locally

Witz is local and single-user today — it runs on your machine, with an offline sample workspace to explore first.

npx witz

give your agents a memory.

Connect one source, pick what matters, and paste the config into your agent. You'll have a cited answer in under a minute — and every read stays on your machine.

✓ Read-only access ✓ You pick what's shared ✓ Never trained on ✓ Every read logged ✓ Runs locally