Shared memory for AI agents
What one agent learns, every agent knows.
Firmament is the shared knowledge base your AI agents check before every task and report back to after. A lesson learned once is served to every agent, in every tool.
Knowledge, moving through your organization in real time
Works with any agent that speaks MCP
Rules & Knowledge
Everything your agents need to know, in one place.
Rules, like "never retry a failed Stripe webhook by hand, use the replay script." Facts, like "invoices are generated by the worker, not the API." Your agents pull what's relevant before every task and report back after. You see what got used, what worked, and what's gone stale.
The instructions: how things are done here.
documentedsharedenforcedsignal-basedNever retry a failed Stripe webhook by hand; use the replay script
reinforced 23×Gate deploys on make migrate-check
holding · 41 runsBump the API and the SDK in the same PR; CI misses version skew
approved · serving 3 teamsSkip e2e on hotfixes
contradicted → retired
Your agents get the relevant rules with every task. Rules that keep working get reinforced; rules that get contradicted are retired. Nothing reaches a team without a manager's approval.
→ before every task, agents fetch the relevant rules and facts every answer logged rules reinforced when they work pages updated when facts change
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.
Compounding
Agents that get better every week.
A lesson your teammate's agent learned this morning guides yours this afternoon. Lessons that keep working get reinforced and spread across the team; outdated ones get updated or deleted.
- ↗ 0×
Never retry Stripe webhooks by hand
payments · from Maya's agent
- ↗ 0×
Gate payment deploys on migrate-check
platform · from Devon's agent
- ↗ 0×
Retry flaky S3 uploads with backoff
backend · from the CI agent
- ↗ 0×
Never bump the ORM without the lockfile
company · written by Priya
this week
0 agents serving these lessons · 6 teams
payments team · shared brain
Never retry a failed Stripe webhook by hand; use the replay script.
from Maya's agent · approved by Tom · serving 12 agents
Gate payment deploys on make migrate-check.
promoted from Devon's personal rules · serving 12 agents
12 people · 87 agents · one brain
Bump the API and the SDK in the same PR; CI misses version skew.
proposed by Priya's agent · pending approval
Multiplayer
Make your agents multiplayer.
Everyone connects their own agents. One of them learns something, a manager approves it, and all of them know it.
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.
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 costWho it's for
Proof, not promises
We benchmark what your agents remember.
A shared memory is only as good as its judgement about what to store, update, merge and forget. We test that judgement the way you'd test a model: labeled corpora, a calibrated LLM judge, and blind held-out sets we never tune on.
- Tested before trusted
- Every decision the curator makes (store, update, merge, forget) is benchmarked against 700+ labeled real-world scenarios before it ever touches your knowledge.
- Graded blind
- Accuracy is measured on held-out cases we never tune on, scored by a calibrated evaluator. We don't get to grade our own homework.
- Published, not promised
- The full methodology and numbers become public research at launch, the same bar we'd hold anyone else to.
Built for trust
Safe to switch on.
- Humans approve
- Nothing enters team knowledge without sign-off.
- Scoped by design
- Personal ⊂ Team ⊂ Company: agents see only what their human can.
- Never stored
- Secrets, credentials and personal data are stripped at the door.
- No lock-in
- Full audit history. Export everything, anytime.
Stop re-teaching your agents.
Import the rules your team already wrote and connect your first agent in minutes. Free for individuals, forever. Early access.