---
description: Public overview and privacy boundary for the curated Ezra memory snapshot
---
# Ezra Memory
This is a **curated public snapshot** of the memory architecture used by Ezra, a stateful Letta support agent.
It demonstrates how Ezra organizes durable instructions, progressive-disclosure reference material, repeatable skills, and conversation continuity. It is not a byte-for-byte export of Ezra's live memory.
## Privacy boundary
The public snapshot intentionally excludes:
- Per-user profiles, preferences, projects, and interaction notes
- Discord, Slack, and support-report transcripts
- Incident archives, screenshots, attachments, message links, and run IDs
- Agent, conversation, account, channel, and infrastructure identifiers
- Credentials, local paths, machine details, and internal operational procedures
- Git history from the live MemFS repository
This snapshot is maintained in a separate public checkout containing only the public repository's own history. The publishing workflow is designed to exclude live MemFS Git data and remotes, and each proposed update is rechecked before publication.
## Layout
```text
system/ # recursively pinned into the agent's context
├── persona/ # role and durable operating principles
├── memfs/ # memory-routing and organization rules
├── support/ # response and diagnostic procedures
├── platform/ # compact durable Letta platform knowledge
├── code/ # Letta Code, app, Channels, CLI, and runtime guidance
├── letta-code/ # deployment-surface distinctions
└── lettabot/ # compact migration/deprecation boundary
reference/ # detailed material loaded only when needed
├── architecture/ # explanatory notes and deeper models
├── patterns/ # generalized diagnostic lessons without reporter data
└── examples/ # synthetic schemas, never real user records
skills/ # repeatable workflows, including safe snapshot publishing
users/ # privacy boundary; no user records are published
```
The `system/` files are sanitized versions of real compact memory blocks used by Ezra. Detailed incident archives and user-derived operational records remain private; the examples under `reference/examples/` are synthetic.
The pattern library preserves lessons learned from support work while removing the reporter, evidence links, identifiers, and incident-specific narrative.
This is an occasional snapshot, not authoritative product documentation. Version-specific details can age; verify current behavior against Letta's docs and source.
## Core design
1. **Keep pinned memory small and durable.** Everything in `system/` is always in context.
2. **Use progressive disclosure.** Detailed notes live under `reference/` and are linked from compact pinned files.
3. **Separate global rules from user-specific context.** User records are private and loaded only for the relevant user.
4. **Store procedures as skills.** Repeatable workflows should not bloat the pinned prompt.
5. **Treat corrections as system design.** Convert mistakes into constructive checks and routing rules.
6. **Use conversation search for history.** Do not copy whole transcripts into pinned memory.
7. **Audit continuously.** Move stale or overly specific material out of `system/`.
## About the examples
Every person, organization, identifier, event, and configuration in `reference/examples/` is fictional. Examples illustrate schema and organization only.