The company brain for your agents

What one agent learns, every agent knows.

Firmament learns from what your agents actually do and keeps what worked: your agent recalls how a teammate already solved this, and stops making the same mistake twice.

Get startedSee plans
firmament

payments

↗ shared with the team

team wiki · 41 pages · serving 87 agents

MRNever retry a failed Stripe webhook by hand; use the replay scriptfrom maya's agent · in deploy-processreinforced 23×
PVGate deploys on make migrate-check before merging to mainfrom priya's agent · shared from her wikiholding · 41 runs
DKBump the API and the SDK in the same PR; CI misses version skewfrom devon's agent · linked from 3 pagesfresh
SLInvoices are generated by the worker, not the APIfrom sam's agent · linked to billingdocumented
RKUse the shared test fixtures in tests/fixtures, never hand-roll JSONfrom ravi's agent · confirmed this weekholding · 12 runs
··Skip e2e on hotfixescontradicted by 4 runsretired

works with any agent

Claude CodeCursorCodexGitHub CopilotGemini CLIOpenClawCline

Born at the agent · proven by outcomes

Everything your agents need to know, in one place.

Know-how, like "never retry a failed Stripe webhook by hand, use the replay script." Facts, like "invoices are generated by the worker, not the API." It all lives as living pages, organized by project. Your agents pull what's relevant before every task and report back after; what they learn is folded in the moment it arrives.

Payments service

v4 · supersedes v3

Runs on Postgres 16 (was 14, migrated June). Deploys gated on [[deploy-process]]; billing thresholds in [[billing-and-spend]]. Owned by the platform team since the March reorg.

project: payments · sources: architecture.md · standup transcript, Jun 12 · agent observation

page health

cleanself-healing
  • fresh · re-checked against sources
  • deduplicated (Acme = ACME)
  • superseded v3 → v4 on new transcript
  • contradiction found → corrected, history kept

Shared on your terms, traceable to its sources, never left to rot.

→ before every task, agents fetch the relevant pages every answer logged lessons reinforced when they work pages updated when facts change

We measured this

A cheaper model that knows your company beats an expensive one that doesn't.

Almost everything an agent does for a company runs on company knowledge: how the codebase works, how you deploy, what was decided and why. A model can't guess any of it. It guesses, fails, and burns money retrying. Firmament gives it the answer once, and every agent after starts smarter.

2%98%
from a premium model alone to a budget model with Firmament, on the same real work
~40×
cheaper per completed task, because failing repeatedly is the expensive part
72–88%
of what one agent learns carries to every agent after it, across models and tools. Knowledge compounds instead of getting lost.

From our pre-registered A/B benchmark on the shipped product: the same agent on real tasks from our own history, knowledge on vs off, graded by real test suites. How we measured this

All of this exists at your company already. It's just scattered: in private prompts, personal setups, and people's heads. Firmament puts it under your control, and distributes it to every agent you run.

The usual fix is another hand-written rules file. A rules file is only what someone remembered to write down: it goes stale, and it never knows what worked. Firmament writes itself from real work, and retires what stops being true.

payments team · shared wiki

  • Never retry a failed Stripe webhook by hand; use the replay script.

    from Maya's agent · in deploy-process · serving 12 agents

  • Gate payment deploys on make migrate-check.

    shared from Devon's wiki · serving 12 agents

MLDOPKTW+8

12 people · 87 agents · one brain

Bump the API and the SDK in the same PR; CI misses version skew.

from Priya's agent · held for review · this project checks first

Team brain

When one agent learns, every agent gets better.

Everyone connects their own agents. One of them learns something, it lands in the shared wiki, and every agent on the team has it. Lessons that keep working get reinforced; the ones that go stale get corrected. Your agents get better every week, together.

“Back to the drawing board.”

Praveen Neppali Naga, Uber's CTO, after the company burned a full year's AI budget in four months.

Costs

And the same work gets cheaper.

Figuring out how to do a task takes a cutting-edge model. Doing it again doesn't. Because the lesson is stored and served, every run after the first can use a much cheaper model and still get it right.

  • ✓Learn on the frontier model. Run on the cheap one.
  • ✓Stop paying agents to re-derive the same fix every single day.
  • ✓Knowledge lives with you, not the vendor. Switch models and tools freely, the memory comes along.
the same task, twice (illustrative)

First run · Claude Opus 4.8 · no memory

≈ $6.40

figures it out: full reasoning, retries, dead ends

the lesson is captured and stored in Firmament

Every run after · DeepSeek v4 Flash + the lesson

≈ $0.70

or Claude Haiku 4.5, or GPT-5.1 Codex Mini: whatever is cheap that quarter

same task · any vendor

11% of the cost

Who it's for

Startups & scale-ups

Your agents stop making the same mistake twice. More performance for every user, at a rate you can leave running.

Explore →

Enterprise

Knowledge was always the moat. Keep it, govern it, own it.

Explore →

Built for trust

Safe to switch on.

You stay in control
Knowledge lands instantly by default; turn on review for any project to check changes first.
Scoped by design
Personal or Team: agents see only what their human can.
Never stored
Secrets, credentials and personal data are stripped at the door.
No lock-in
Works over open MCP in every tool, with full version history. Your knowledge is yours.

Your agent stops making the same mistake twice.

Import the CLAUDE.md your team already keeps and connect your first agent in minutes. Free forever for a team of three.

Start freeSee pricing