Format Specification · v1.0.0

The .mem Format

A structured text format for encoding enterprise software environments as self-contained, human-readable memory files.

What is .mem?

The .mem format is a plain-text document structure designed to encode complex, multi-faceted knowledge about enterprise software environments into portable files. Each .mem file is self-contained — it holds everything needed to understand a single company's engineering context, including its architecture, operational data, behavioral rules, and narrative history.

The format draws inspiration from structured document standards like YAML and Markdown, combining human readability with machine-parseable sections. Every SWE-Enterprise world includes a .mem file as its canonical reference document.

Format Specification

A .mem file is organized into sections, each prefixed with a !section_name directive. Sections are ordered by convention but parsed independently.

!meta

Document metadata: name, version, description, and source attribution. Every .mem file begins with a !meta block that defines its identity. This is the only required section.

!state

Current session state: task description, completion status, decisions made, and next steps. Tracks the active work context and progress through objectives.

!arch

Architecture documentation: system design, component relationships, data flow, and technical decisions. Mirrors an Architecture Decision Record (ADR) in structured form.

!data

Structured data: domain distributions, sample counts, file locations, and quantitative metrics. Supports key-value pairs and nested structures.

!rules

Behavioral rules and best practices: coding standards, operational guidelines, preference constraints, and invariants that apply to the described environment.

!mem

Narrative memory entries: chronological records of decisions, events, and context changes. Each entry is timestamped and describes a discrete unit of work or discovery.

Why .mem?

Human-Readable

Plain text with transparent structure. No binary formats, no opaque encodings. Open and inspect any .mem file in any text editor.

Git-Friendly

Line-based format enables meaningful diffs, clean merges, and standard version control workflows. Every change is traceable.

Self-Documenting

Each file contains its own context, metadata, and history. No external database, no API, no runtime dependencies required to read it.

Extensible

Custom sections can be added with the !name convention. The parser ignores unknown sections, making the format forward-compatible.

Tool-Agnostic

No proprietary tooling required. Any language can parse .mem files with a simple regex or line-by-line reader. Importable anywhere.

Proven at Scale

Powering 24,316 enterprise worlds with 97,290 extracted evaluation tasks. Format is battle-tested at production scale.

Example

!format 1.0.0
!meta
  name = FinTech Payment Platform
  description = Payment processing microservices company
  domain = fintech

!arch
  Core Architecture & Concepts
    - Event-driven payment pipeline (Kafka)
    - Three microservices: gateway, processor, ledger
    - Multi-region deployment for compliance

!data
  employees = 142
  culture = Remote-first
  stack = Node.js, PostgreSQL, Kafka, Kubernetes
  incidents = 14 logged · 2 critical

!rules
  - All payment mutations require audit trail
  - Currency conversion must round down
  - Refunds over $50K require manual approval

!mem 2026-01-15
  feat: initial payment gateway service
  - Created payment intake endpoint
  - Added webhook signature verification

Usage in SWE-Enterprise

Every world in the SWE-Enterprise dataset includes a canonical .mem file that describes the company's full engineering context. These files serve as:

Ground Truth Reference

The .mem file contains all structured metadata used for grounding evaluation tasks. RAG systems can be benchmarked against these canonical records.

Context Window

Training agents can load the .mem file as a compressed context window — understanding the company structure before interacting with its codebase.

Audit Trail

Chronological !mem entries track the narrative of each world, providing temporal context for mission sequences and architecture evolution.

Cross-World Consistency

Standardized format ensures every world follows the same schema, enabling consistent parsing, querying, and evaluation across all 24,316 worlds.

Editions

From free samples to enterprise-wide deployment. All editions include full Commercial Training License.

Feature Evaluation Research Professional Enterprise
5 Preview Worlds
1,000 Worlds
5,000 Worlds
Full 24,316 Worlds
Commercial License
QA Report
Dataset Updates 6 months 12 months Ongoing
Priority Support
Evaluation
5 worlds · Evaluate quality before you commit.
  • 5 enterprise worlds
  • Sample from all 10 domains
  • All 5 components per world
  • Unrestricted evaluation
Request Quote
Professional
5,000 worlds · Built for commercial AI teams.
  • 5,000 enterprise worlds
  • Stratified domain sampling
  • 20,000+ RAG evaluation tasks
  • Full commercial training license
  • Priority support
Request Quote