Prove the loop: connect one tool, ingest a real source, and watch agents recall the same facts.
- Ingest
- 25 MB / month
- Recalls
- 500 / day
- Members
- 3
- Connectors
- 1
- MCP retain, remember, and recall
- Console Sources upload
- One showcase connector
One memory. Every AI tool.
Shared organizational memory for Cursor, Claude, ChatGPT, and Codex — so agents recall the same facts with sources, instead of inventing a new company each session.
The problem
Each chat starts empty. People paste the same policy, owner, and “we decided this last month” into Cursor, Claude, and ChatGPT. Answers drift. Two agents give two versions of the firm. Nobody can show where a claim came from.
Keystone is where that knowledge actually lives: facts with sources, recalled over MCP by every connected tool. When sources disagree, rival claims are scored by a typed decision model — conflicts stay visible instead of being silently overwritten.
Why it's worth it
Built for teams that already use AI tools and are tired of re-teaching the company — not another prompt pack or chat-only memory that quietly invents resolutions.
Policies, owners, and decisions live once. Agents recall what is current — with citations — instead of last week’s chat residue.
Cursor, Claude, ChatGPT, and Codex read one workspace memory over MCP. Connect once; the firm’s record does not fork by product.
When sources clash, TypeSafe models score the rivalry. Real conflicts open an inbox. Nothing important is silently overwritten.
Retained statements point at a document, note, or connector payload. You can show where a claim came from — and when it was true.
Jev returns duplicate, extends, contradicts, or unrelated — not free-form chat. Faster and cheaper than asking GPT to invent a resolution.
Members, scopes, and connectors scale with the firm. Free proves the loop; Team is when memory is shared infrastructure.
What sets us apart
Most memory products ask a generative LLM to “resolve” contradictions. That is slow, expensive, and free to invent an answer. Keystone routes rival claims through TypeSafe System One models (Jev) — bounded choices with calibrated probabilities: duplicate, extends, contradicts, or unrelated.
List-like facts can coexist. Exclusive clashes open an operator inbox. Confidence below threshold never silently overwrites. That is governance memory — the reason teams pay once the record is load-bearing.
subject · attribute · existing claim · new claim
Powered by TypeSafe Jev on the adjudicator path — structured decisions software can trust.
How it works
Point Cursor or another MCP client at your workspace. Free includes one connector so the ingest path is real, not a screenshot.
Upload sources, retain through MCP, or sync a connector. Extraction turns prose into facts; TypeSafe Jev adjudicates rival claims with typed confidence.
Ask from any connected tool. Keystone returns what is current, flags what is contested, and meters usage against your plan.
Pricing
Free is enough to connect a tool, ingest a real source, and prove shared recall. Team is for when every agent in the firm should read the same record — more seats, connectors, and the conflict inbox that keeps disagreements honest. Caps are enforced: a call past a limit is refused, not silently billed.
Prove the loop: connect one tool, ingest a real source, and watch agents recall the same facts.
When memory is infrastructure: every seat and AI tool reads one record, with conflict inbox and room to grow.
Contracted limits, identity, and onboarding when the firm cannot run on a shared cap.
Need more memory or recalls? Upgrade or contact us. Self-serve overage is not billed yet — a call past a limit is refused until the period resets.
Learn more: What is organizational memory? · vs Mem0 · vs Zep