Decisions
graven in stone.

Not lost in chat.

Decision memory for AI-assisted projects and backend services — every choice stored with rationale, linked by dependencies, versioned, and available through REST or MCP.

chat — acme/backend
You

Let’s not use Redis. LRU in-memory, 5-min TTL. We don’t need the complexity.

Agent

Agreed. LRU with Map-based approach, 5-min TTL.

· · · 130 lines later · · ·

Agent

Setting up Redis with ioredis and a connection pool

This happens every session. Every context switch. Every new chat.

Decision graph

Every decision connects.

Internal engine extracts relationships and assigns confidence weights to every edge. The graph returns not just what depends on a decision, but how strongly.

Change a decision — the agent sees ranked downstream impact before a single line of code moves.

decision · confirmedAuth: session-baseddecision · confirmedCaching: LRUdecisionRedis: rejectedconstraintToken rotation: 15massumptionCache-Control headersdecisionCDN: origin-onlylesson learnedIn-proc store OK⚡ blast radius: 2 affected

Code is the consequence. Decisions are the cause.

What changes

The agent stops guessing.

Closed questions stay closed.

The agent works inside the lines your team already set. Decisions, assumptions, constraints — one system.

See impact before the change.

Blast radius with weighted dependencies — exact list of affected entities, severity, depth. One second.

New chat, full context.

Agent opens a fresh session and instantly knows every decision, constraint, and where you left off.

Assumptions that know when they break.

Every assumption carries invalidation triggers. When the world changes, the system flags it.

The graph builds itself.

You write decisions. AI enrichment finds relationships — dependencies, derivations, potential tensions. Automatically.

Existing projects welcome.

Built-in onboarding extracts decisions from your codebase, configs, and git history. Works on day one.

In practice

How it works in real projects.

340decisions

B2B SaaS platform

Team of 6 + 3 AI agents, 14 months in prod

Migrating to GraphQL? 89 entities affected. We saw the blast radius before writing a single resolver.

580decisions

Fintech compliance engine

2 founders, AI-first, 47 regulatory constraints

Regulator changed KYC rules. "What depends on identity verification?" — 142 entities, 12 critical. Scoped in minutes.

207decisions

AI-native dev tool

Solo founder with Cursor + Claude Code

One tech stack decision affected 86% of the project. One assumption had 185 decisions depending on it.

Who is this for

Built for how you actually work.

Founder / PM with 30+ decisions

You don’t remember all the dependencies anymore. Your agent remembers even less. Graven holds the map — so neither of you has to.

Team where multiple people (and agents) decide

Everyone makes decisions that affect others. Without the dependency graph, conflicts surface on production. With it — they surface before code.

Any project longer than one session

100 messages is enough for an agent to forget a decision from the start. After 3 sessions — it doesn’t exist. After 10 — the team re-makes it from scratch.

Stop re-explaining.
Start building.