lema · docs
Introduction
Models are interchangeable. Your reasoning isn’t.
AI models are becoming a commodity — this quarter’s best is next quarter’s default. Your team’s judgment isn’t. What you decided, what you ruled out, and why it still holds is the asset that compounds — and today it evaporates into PR threads nobody re-reads. lema keeps it as a queryable, cited record of the why.
The why, made queryable
Ask whyand get a cited answer from your team’s own decision graph — the rationale, not just the record.
Agents write the code. lema keeps the why — the alternatives ruled out, the constraints that still apply — so it survives turnover.
The expertise lives in your graph, not the model’s weights — served to any agent over MCP. Switch models without losing what your team learned.
lema does the labor — synthesize, cite, ground, draft. It never makes the call. Judgment stays human; the record makes it reusable.
What your agent gets
Tight, typed claims — each carrying its source, summarized not quoted, never whole documents. The example below is the answer to “why did React move from classes to Hooks — and what was ruled out?”
- chosen· reactjs/rfcs#68
Hooks let function components hold state and lifecycle, so stateful logic can be reused without wrapper components.
- rejected· reactjs/rfcs#68
A mixins-style sharing mechanism was ruled out before Hooks — name collisions and implicit dependencies.
- constraint· rules of hooks
Hooks are called at the top level in the same order each render — never inside conditionals or loops.
- consequence· reactjs/rfcs#68
Classes stay supported; Hooks were additive and opt-in, not a migration mandate.
The rejectedclaim — the why-not — is the part no other tool surfaces: an agent learns what was already argued and ruled out, so it doesn’t reopen it.
When it stays silent
When the record holds no decision on a question, lema says so. That is unknown — never approval. An honest gap beats a confident guess, so the product surfaces the gap rather than dressing one up as an answer.
How this doc is organized
- QuickstartWire a repo and ask the graph.
- Connect your repoSign in, mint a key, and serve your own decision graph to every agent session over hosted MCP.
- The why, not the whatWhat a typed claim is, and why it always carries its source.
- FusionCheck whether an upstream project already argued an approach and ruled it out.
- ToolsThe full set of tools an agent sees.
- API referenceThe same graph over plain HTTPS — ask, retrieve, check, and account endpoints with a lema_live_ key.
- VerifyThe CI check on every pull request: did the diff do what the description says?
- Public recordShareable, no-signup URLs for public decisions and repo records — plus the no-auth read API.
- EvaluationHow faithful lema stayed to the source on repos it had never seen — misses included.
Record the why once, and every future session — human or agent — arrives grounded.