Decision layer for AI coding agents
The engineering decision layer for AI coding.

Capture the engineering decisions that matter. Replay them across every AI coding session.

Stop re-explaining. Stop losing key decisions. Stop wasting tokens. Stop architecture drift.Keep every AI coding assistant and project aligned — and increase AI ROI.

No account required · Decisions stored locallyWorking in a team? See Contexer for Teams →

Supported agents

  • /claude code
  • /cursor
  • /codex
  • /gemini cli
· The real problem

Your AI doesn't forget code.
It forgets why your team made certain decisions.

01

Architecture drift

The AI proposes solutions that violate architectural decisions your team made months ago - quietly, in code that compiles.

02

Tribal knowledge

Critical engineering knowledge lives inside senior engineers' heads instead of being available in every AI session.

03

Cross-agent inconsistency

Claude Code, Cursor, Codex, and Gemini CLI all produce different implementations because each one starts every session with different assumptions.

04

AI re-explanation tax

Every new coding session begins with re-explaining the same conventions, constraints, and architecture you've already approved.

· Workflow preview

Capture, propose, approve. Right inside your terminal.

These mockups show the intended flow. The actual UI may differ. The idea is decisions captured, reviewed, and replayed where your AI coding assistant already lives.

01Capture
Live
>
Captured as global convention
Rule
Release sequence: branch → bump → PR → merge → tag
Loaded into 0 sessions · applies to all future repos

State a rule. Contexer stores it.

Directives you state outright become global conventions, loaded into every future session.

02Propose
Idle
Pending decisionAI proposed
Use RabbitMQ for async jobs, not Kafka.
Rationale: lower ops overhead, fits current scale.
1 approve
2 edit
3 ignore

AI surfaces a decision mid-work.

Architecture and constraint decisions are parked for review, never replayed until you approve.

03Approve
Idle
>
Approved · v1
Decision
Use RabbitMQ for async jobs.
immutable
replayed across agents

One keystroke. Versioned forever.

Approve, edit, or ignore. Approved decisions are immutable and replayed across every agent.

· Benchmark · 540 sessions

Same model. Same repo. Same question.

We ran Claude Sonnet 5 and Opus 4.8 across six memory conditions, scored by code — not a judge model. Here's one question, two runs.

No memoryBare session
> why did we choose Postgres over MySQL?
· grep -r "postgres" .
· reading src/db/config.ts
· reading docker-compose.yml
· searching git log
· reading README.md
· scanning ADRs directory (none)
"I can't determine the reason from the codebase alone."
6
turns
165k
tokens
$0.11
cost
With ContexerInjected
> why did we choose Postgres over MySQL?
"Postgres was chosen for JSONB support and stronger transactional guarantees. MySQL was rejected — see decision arch-014."
Cited
arch-014 · approved 2024-11-03 · 3 evidence links
1
turn
33k
tokens
$0.04
cost
See full benchmark →

540 sessions · 2 models · 6 conditions · deterministic scoring · independent validator · adversarial review · raw JSONL published

· What Contexer actually captures

Engineering decisions. Not chat history.

Architecture Decisions

"We standardized on PostgreSQL."

Constraints

"Never expose S3 buckets publicly."

Conventions

"Terraform modules live under infrastructure/modules."

Patterns

"Every API emits OpenTelemetry traces."

These are engineering decisions - not chat history, not transcripts, not session logs.

· More than agent memory

AI agent memory remembers conversations. Contexer remembers the engineering decisions you approved.

CLAUDE.md, AGENTS.md, .cursorrules, and per-assistant memories capture whatever one person remembers to write down, and get overwritten with no review. Contexer captures typed engineering decisions, holds the risky ones for your approval, and versions them so the latest approved revision is what every agent sees.

Traditional AI memory
Contexer
What it stores
Conversations and session history
Typed engineering decisions with rationale
Review
No review - stored as-is
AI-proposed architecture/constraint decisions and updates held for your approval
Versioning
Files overwritten in place
Approved decisions are immutable; changes create a new revision with full history
Replay
Whatever the file says, every time
Latest approved revision replayed every session
Awareness
Session-level
Cross-session, cross-repository
Coverage
One AI assistant at a time
Claude Code, Cursor, Codex, and Gemini CLI
· Pricing

Start free. Upgrade when your team grows.

Most popular

Free

Freeforever

The same core product, available two ways. Use the local CLI and JSON store with no account, or sign in for managed cloud sync and the web dashboard.

MIT licensed · No account required for local OSS

Open Source CLI

Local JSON store. Fully offline. No web app or cloud. No account required.

Personal Cloud

Encrypted sync, web dashboard, and cross-machine replay.

  • Capture architecture, constraint, convention, and pattern decisions
  • AI-proposed decisions held for your review
  • Approve, edit, skip, or dismiss before replay
  • Injected into every Claude Code, Cursor, Codex, and Gemini CLI session
  • Web dashboard to review history and manage decisions (Cloud)
  • One-click upgrade to Teams

Teams

$10per repository / team / month

One repository per team. Invite members and share approved engineering decisions across everyone's AI coding sessions.

30-day free trial · No credit card required

  • Everything in Free Cloud
  • Unlimited team members with OAuth sign-in
  • One shared repository per team
  • Shared decision repository via team MCP endpoint
  • Role-based approval workflow (lead / member)
  • Confidence scores with supporting evidence - leads only
  • AI-generated canonical suggestions and updates
Contact us

Organization

Custom

For orgs running multiple teams. Cross-team sharing, enforcement, CI/CD drift detection, SSO, and audit trail.

SSO · Audit trail · Pipeline drift detection

  • Everything in Teams
  • Cross-team decision sharing across the org
  • Enforcement policies (require approval, block on drift)
  • CI/CD pipeline hook - detects and rates architecture drift on every PR
  • SSO login (SAML, OIDC)
  • Full audit trail and access controls
  • Dedicated support and onboarding
· Quick start

Two commands. Under two minutes.

Requires Python 3.12+ and uv. Auto-detects Claude Code (~/.claude), Cursor (~/.cursor), Codex (~/.codex), and Gemini CLI (~/.gemini) and wires them in.

bash · install
# Step 1. install
$ uv tool install contexer

# Step 2. wire into your AI assistant
$ contexer install

-> detected ~/.claude (Claude Code)
-> detected ~/.cursor (Cursor)
-> detected ~/.codex (Codex)
-> detected ~/.gemini (Gemini CLI)
-> wired all four. restart your AI assistant.

Restart your AI assistant and open any git repo. Contexer runs from there.

· Cost

~26 tokens per rule. Paid once per session.

Only constraints and conventions are replayed every session. Architecture and pattern decisions cost nothing until the work actually needs them. Store lookups are sub-millisecond and run before the response is generated.

Pre-loaded rules
Tokens at session start
5 rules
~125
10 rules
~250
25 rules
~625

Deduplication is deterministic - >70% token overlap is dropped. No LLM call. No tokens.

session start context
## Project rules - apply to ALL tasks in this repo:
- [convention] Use uv, not pip, for
  all dependency management
- [constraint] Never commit untested
  code - CI blocks merges below coverage

2 architecture decision(s) stored. Call
get_context before reading files for
questions about architecture or rationale.