Skip to content
Mithrandir
Talk with founders
Talk with founders

Your company's memory, inside every AI agent your people use.

Mithrandir learns from your company's history and keeps learning as you work: every decision, with the reason behind it, from a sales call to a line of code. Each person's own agent gets the answer, with only what that person is allowed to see.

Try it on a real codebaseSee how it's different

Not a search box, and not a chatbot over your documents: a memory that knows why.

A simulated session at a made-up software company. Acme, named across this page, is one of its customers. Pick a question or type your own, then ask as the Intern to see what gets withheld. Try the live demo

A new account manager sees Acme paying 20% under list and wonders if it is a mistake.

Claude· director clearance

vs intern · adds the transcript line, the decision, the quote and the contract amendment

  1. It is a decision, not a typo. Ayşe K., Head of Sales, made the call in the 3 September sales sync: 20% through Q4, on the condition that Acme signs the annual plan.

  2. It applies to Acme Corp's annual plan, which renews in December 2026. The account team is Ayşe and Deniz.

  3. Sales sent the revised quote the same afternoon, and Legal wrote the discount into the contract as amendment 2 the next day.

  4. Engineering shipped percentage discount codes on 5 September so billing could apply it. That code exists because of one sentence in a meeting.

Works with

  • Claude
  • Claude Code
  • ChatGPT
  • Codex
  • Gemini
  • Antigravity
  • Cursor
  • Hermes Agent
  • Windsurf
  • GitHub Copilot
  • Cline
  • Kiro
  • Devin
  • OpenCode
  • Roo Code
  • Goose
  • Warp
  • JetBrains AI
  • Replit
  • Trae

Search finds files. RAG finds paragraphs. Mithrandir knows the decision.

Ask the same question of the three kinds of AI knowledge tool a company can buy, and you get three different things back.

Why does Acme pay on net-45 instead of net-30?

Enterprise search

Finds documents that mention it

  • Acme MSA 2024.pdf
  • Re: Acme payment terms · 41 messages
  • Q3 billing changes.xlsx
  • Finance sync notes, Aug 2026
  • + 23 more results

Everything that matches the words. Reading it, and working out which file still holds, is left to you.

AI chat over your documents (RAG)

Finds the paragraph that looks closest

Acme's payment terms are net-30, as set out in clause 7.2 of the Master Services Agreement.

source: Acme MSA 2024.pdf, p. 6

Fluent, confident and out of date. The best-matching paragraph wins, whether or not it still holds, and nothing in it says why.

Mithrandir

Knows the decision and why it was made

Net-45 since August 2026. Acme's procurement pays every vendor on a 45-day cycle and would not sign net-30; Finance agreed because the annual plan is prepaid in full.

It replaced the net-30 in the 2024 agreement, which still applies to invoices issued before then.

  • decision · #legal-sales · aug 2026
  • agreement · Acme MSA 7.2, as amended

The answer that holds today, the reason for it and what it replaced, each tied to its record, and only what the person asking is cleared to see.

Illustrative answers, on the same made-up company as the demo above; Acme is one of its customers.

How this compares to Glean, Copilot and Unblocked

Measured, not estimated

An eighth of the cost. Every answer sourced.

8.3 times
lower cost$0.031 vs. $0.234 per output at Claude Opus list prices, against an agent mirroring the request with the exact same tools, minus Mithrandir. An 88% smaller AI bill, for better results.
100 percent
of answers grounded in external sourcesNot a single output returned without exact sources out of hundreds of trials.
zero
leaks of contextOut of hundreds of trials and thousands of customer entries, not a single leak of context.

Cost, tokens and speed: the same questions on our public demo, answered by an agent using Mithrandir and by a coding agent with the same model and the same repository, without it. Medians of per-question ratios; dollars at Claude Opus list prices. Quality and sourcing come from our pilot.

All six numbers and how we measured

It reads where the decisions already happen.

Slack, Linear and Jira, GitHub, Google Workspace, Notion, meeting transcripts, your custom tools...
Anything you can think of.

Nothing bothers twice.

  1. 1

    Connect your sources.

  2. 2

    Mithrandir builds the memory.

  3. 3

    Every agent your people already use can ask it.

See what it builds from them

Security your CISO can verify, not hope for

Most enterprise AI filters answers, not access. Mithrandir is built the other way around.

What a security review will find

Nothing to leak

Restricted records never enter the model's context. No prompt rule to argue with, nothing for a jailbreak to pull out.

Complete for each clearance

Intern and director ask the same question and each gets a full answer for their level. No errors, no redactions, no leaks.

Your data stays yours

Self-hosted in your own cloud, on a model you can host yourself. Nothing trains on your data.

Is Acme's 20% discount a mistake?

  • Sales sync · transcript
  • SAL-214 · Send Acme the revised annual-plan quote (20% through Q4)
  • Acme MSA · amendment 2
  • e7b21c4 · billing: percentage discount codes for contracted discounts (PR #418)
  • Acme Corp · customer record

5 records in the director's answer. The 3 marked are beyond an intern's clearance.

One price per person. The whole company.

Early-customer price

$29Standard price $39per seat, per month

The standard price will be $39. Until then, every team pays the same $29, from sales to engineering.

Lock in $29 for as long as you stay.

Your team50 people

$1,450 per month for 50 people, $500 a month under the standard price.

Book a call

Thirty minutes with the founders.

Included in every seat

  • The whole company: sales, legal, finance, people and engineering
  • Every connected source and every agent your people use
  • Answers limited to what each person is cleared to see
  • Self-hosted, in your own cloud or on your own hardware
  • Tamper-evident audit log you can stream to your SIEM

Not included: the model Mithrandir runs behind the scenes. Use your own provider account, a model you host, or have us run it for you on AWS. Estimate how much smaller your AI bill gets

Backed by ongoing research at Carnegie Mellon University, School of Computer Science, Language Technologies Institute, on the limits of rule-following in language models.

Rememberwhy.

Securing the why is the key to the gold buried under the mountain.

Remember more. Spend less.

Thirty minutes with the founders.