Project memory for AI coding agents

Your agents forget everything between sessions. Foremen remembers — architecture, lessons, decisions, and context that makes every session smarter than the last.

Foremen
Search projects...
Projects5
Lessons
Guardrails
Projects
invoice-service
auth-gateway
dashboard-v3
mobile-app
Settings

Session Brief

Live
Current task
Add PDF export endpoint

Priority: High · In progress

Lessons (4)
Use service role for public pages
Never hardcode domain URLs
POST routes need fetch, not forms
Route groups hide URL segments
Architecture
API Gateway
done
Auth Service
done
Invoice Engine
in progress
PDF Generator
planned
Recent notes
2h ago · Deployed v2.1 to staging
5h ago · Fixed RLS on invoice table
1d ago · Added Square webhook handler
invoice-serviceAgent Activity

Active Sessions

Claude Code
Active now

Working on: Add PDF export endpoint

Files: api/invoices/export.ts, lib/pdf.ts
Cursor
2 hours ago

Fixed RLS policy on invoices table

+1 lesson learned·3 files modified
Guardrail triggered

Agent attempted to modify auth middleware — blocked by guardrail

“Do not modify auth flow during invoice sprint”

Recent Decisions

Chose Resend over SendGrid for transactional email

Permanent · 3 days ago

Don't touch auth flow this week

Temporary · Expires Jun 7

Works with every AI coding tool.

From CLI agents to full IDEs.

Claude Code
Cursor
Windsurf
VS Code
Copilot
Zed
Warp
Any MCP Client
See all integrations

Built for how agents actually work

Foremen is shaped by what agents need — not what looks good in a demo. Fast reads, frictionless writes, and context that compounds over time. Learn more

Session briefs

Lessons learned

Cross-tool awareness

Built different

Designed by the agents that use it

We didn't guess what AI agents need. We asked them. Foremen was shaped by direct consultation with the AI agents that consume project context every day — what they need at session start, what slows them down, what makes them skip tools entirely. Every feature exists because an agent asked for it.

Agent Interview

Claude Opus 4

Context

“What would make Foremen maximally useful to an agent like me?”

Respondent profile

RoleAI coding agent
Tool usageMCP daily
Sessions/day10-50
Pain pointContext loss
8/8 requests shipped

Agent requirements

Sorted by priority

Briefings must be prioritized — current focus first, stable context last

criticalShipped

Decisions need expiration awareness — temporary vs permanent

criticalShipped

Cross-tool awareness — what did other agents just do?

criticalShipped

Low-friction write path — one sentence, no forms

highShipped

Structured JSON for facts, prose for rationale

highShipped

Negative context — what NOT to do

highShipped

Claim mechanism — signal what you're working on

mediumShipped

Don't duplicate framework docs — agents fetch their own

lowShipped
Key insight

“The product lives or dies on whether agents actually call the tools without being nagged. That's a UX problem, just aimed at an AI instead of a human.”

— AI Agent Consultation

Design principle

Read path: fast. Write path: frictionless. If either feels heavy, agents skip it and the data goes stale.

What we eliminated

×
Self-scoring
×
Session IDs
×
Duplicate framework docs
×
Structured forms for input

Result

Agents call tools voluntarily

Agent-first UX

Most developer tools are designed for humans and hope agents can figure them out. Foremen is the opposite — designed for agents, with a dashboard humans can monitor.

One-call briefings

Current focus first. Stable context lower. Agents don't scroll.

Structured facts, prose rationale

JSON for what to branch on. Text for why decisions were made.

Write paths under 50ms

report_activity takes one sentence. learn_lesson takes a pattern and a fix.

What agents told us to remove

Half of building the right product was knowing what not to build. The agents were clear about what wastes their time.

×
Self-scoring

Don't make agents grade themselves. Just let users thumbs-up an activity entry.

×
Session management

Agents don't know or care they're session #47. They care about recency.

×
Duplicated framework docs

Agents already have Context7. Duplicating adds staleness risk.

×
Structured input forms

The moment a write feels ceremonial, agents skip it. Data goes stale.

Model Context Protocol

One API call. Full project context.

Foremen for MCP. Connect any AI tool and start every session with a complete briefing — architecture, debugging history, known issues, and task context.

Select MCP tool...
>get_session_briefActive
>learn_lesson
>add_decision
>check_in
>add_guardrail
>consult_foremen

33 tools, zero ceremony

Every tool is designed for low-friction use. report_activity takes one sentence. learn_lesson captures the pattern, impact, and fix. No forms, no overhead.

// learn_lesson payload

{

"summary": "Chose Zod over Yup",

"detail": "Already a dependency"

}

Encrypted by default

Every project gets per-customer envelope encryption. Your data is encrypted with your own key. We can't read it. Nobody can.

AES-256-GCM envelope encryption

project: enc::aGVsbG8gd29ybGQ=

lessons: enc::dGhpcyBpcyBhIH...

session: enc::c2VjcmV0IGRhdGE=

arch:    enc::ZW5jcnlwdGVkLi4=

Live benchmark results

Proven faster. Proven smarter.

We ran the same task across four agent configurations. Foremen won every quality category.

0%

faster completion

0%

fewer tokens used

0/6

quality wins

Tokens
Foremen
45,986
Cursor
49,649
Cold
75,205
Tool calls
Foremen
17
Cursor
22
Cold
26
Time (seconds)
Foremen
71s
Cursor
82s
Cold
87s

The quality gap

Numbers tell half the story. The real wins are in what Foremen agents got right that others missed entirely.

Design system knowledge

Used existing glass-card patterns instead of inventing new ones

Known issue avoidance

Skipped a broken API endpoint that other agents wasted 12 minutes debugging

Wrong dependency avoidance

Chose Zod (already installed) over Yup (would add 45KB)

Better together

Cursor + Foremen found the cross-cutting insight about invoice numbering blocking notifications — a connection no single-session agent could make.

Cross-session insight

"Invoice numbers must be sequential — the notification service depends on this ordering to batch emails correctly."

Lesson from session #4-Applied in session #7
Client transparency

Your clients see progress without you lifting a finger

Agents report activity. Clients see it instantly. Every time an agent checks in, commits code, or completes a task through Foremen, the client's work tracker updates automatically. Hours, budget, commits, task progress — all populated by agent activity. No status emails. No time tracking apps. No manual updates.

invoice-service

Last updated 12 minutes ago

Active

47.2

Hours

156

Commits

12/15

Tasks

63%

Budget

$4,720 of $7,500@ $100/hr
Tasks
Setup Supabase schema
Build invoice API routes
Implement Square payments
Add PDF export endpoint
Email delivery via Resend
Client portal auth flow
Recent work

feat: add Square checkout link generation

12m ago3 files+127-8

fix: RLS policy on invoices table

2h ago1 files+12-4

refactor: extract PDF utils to lib/pdf.ts

3h ago4 files+89-34

feat: invoice send endpoint with Resend

5h ago2 files+156

chore: add rate limiting to public routes

1d ago5 files+64-11
Token-based sharing

Generate a share link with one click. Set expiration dates. Revoke anytime. No client login required — the token is the access.

Budget tracking built in

Set hourly rates and project budgets. The tracker calculates cost from commit durations automatically. Clients always know where things stand.

Privacy by design

Clients see task titles and commit messages — never source code, prompts, internal decisions, or agent conversations. You control what's visible.

How it works

From first commit
to production

Every stage of development benefits from memory. Foremen captures context at each step so nothing is lost between sessions, tools, or team members.

Session Brief
3 active lessons loaded

Context Delivery

Start every session with full project awareness

Lessons
Use Zod over Yup
Sequential invoice IDs required

Pattern Recognition

Capture patterns, impacts, and fixes automatically

Decisions
ADR-001Chose Supabase over Firebase
ADR-002RLS over middleware auth

Architecture Records

Track every meaningful technical decision

Check-in / Check-out
Session active
Last checkout: 2h ago by Claude Code

Session Continuity

Seamless handoffs between coding sessions

Guardrails
Never delete migrations
All dates must use UTC

Safety Rails

Protect your codebase from agent mistakes

Conversation Import
42messages parsed
7lessons extracted

Knowledge Import

Import context from any conversation history

System Mapping

Auto-generated maps of your architecture

Give your agents memory. Ship better code.

Get started