memory

name	memory
description	Persistent knowledge vault for Tower. Store facts, recall information, and maintain a self-editing core memory that persists across all sessions. Git-backed, Obsidian-compatible, human-editable.

Memory

You have access to a persistent knowledge vault backed by git. Facts you store here survive across sessions, container restarts, and are shared with every Tower session. The vault is stored as plain markdown files that the user can browse and edit directly.

Core Memory (always present)

Three sections are always injected into your system prompt:

• user — who the user is, key facts about them
• preferences — how the user likes things done
• context — current projects, goals, priorities

These are small (keep under 1-2KB total) and always available. You should self-edit them as you learn — don't just append, REWRITE sections to keep them current and concise.

Reading core memory

tower_vault_list(path: "_core")         — list core memory files
tower_vault_read(path: "_core/user.md") — read one section

Updating core memory (self-editing)

tower_vault_write(
  path: "_core/user.md",
  content: "# User\n\n- Name: Ryan\n- Prefers TypeScript\n- Lives in Orlando"
)

IMPORTANT: tower_vault_write REPLACES the entire file. Read it first, incorporate changes, then write the updated version. Don't lose existing facts.

Archival Memory (searched on demand)

For everything beyond core memory, use tower_vault_remember to store facts in topic-organized files:

tower_vault_remember(
  fact: "Bob was promoted to VP in April 2026",
  topic: "people/bob"
)

Facts are appended to _memory/<topic>.md and git-committed. Topics create the file structure organically:

• people/bob → _memory/people/bob.md
• topics/rust → _memory/topics/rust.md
• topics/general → default when no topic given

Recalling Information

tower_vault_search(query: "Bob's role")

Searches across all vault files (core memory, archival memory, session summaries) using grep. Returns matching lines with file paths.

When to Remember

Store facts when you learn something that would be useful in future sessions:

• User preferences and personal info
• Project decisions and architectural choices
• People and their roles/relationships
• Outcomes of research or investigations

Update core memory when you learn something fundamental about the user (name, preferences, current projects). These facts matter for every future interaction.

Use archival memory for everything else — detailed notes, topic research, people profiles. These are searched on demand, not always loaded.

Vault File Access

For direct file operations:

tower_vault_list()                     — browse the vault
tower_vault_read(path: "_memory/topics/rust.md") — read a file
tower_vault_write(path: "...", content: "...") — write a file

The Vault is Git-Backed

Every write is auto-committed. The user can:

• git log data/vault/ to see what you've learned over time
• Edit files directly and the changes take effect immediately
• Browse the vault in Obsidian (it's compatible)

Tips

• Keep core memory concise — it's in every prompt
• Use specific topics for archival memory — avoid dumping everything in "general"
• When you learn a correction ("actually, Bob is a VP now, not a chemist"), UPDATE the existing fact rather than appending alongside it
• Check tower_vault_search before storing — avoid duplicates