Memory

Adjutant has persistent long-term memory stored as plain Markdown files. The memory system lets Adjutant remember corrections, decisions, people, project knowledge, and patterns -- and recall them later.

How It Works

Memory is stored under memory/ in the Adjutant directory. Each category has its own file. When you tell Adjutant to remember something, it auto-classifies the content and appends it to the right file with a timestamp.

The agent loads memory/memory.md on demand when a conversation touches past decisions, corrections, preferences, or other memory-backed topics. Individual memory files are then loaded as needed.

Directory Structure

memory/
├── memory.md                    # Index -- loaded at startup
├── facts/
│   ├── people.md                # People, roles, preferences
│   ├── projects.md              # Learned project knowledge
│   ├── decisions.md             # Decisions and rationale
│   └── corrections.md           # Mistakes and corrections
├── patterns/
│   ├── preferences.md           # Communication/style preferences
│   ├── workflows.md             # Recurring workflows
│   └── exceptions.md            # Edge cases and gotchas
├── summaries/
│   ├── weekly/                  # Weekly journal digests
│   └── monthly/                 # Monthly summaries
├── conversations/               # Notable conversation takeaways
└── working/                     # Ephemeral working memory (auto-cleaned)

Commands

Telegram

Command	What it does
/remember <text>	Save something to memory (auto-classified)
/forget <topic>	Archive memories matching the topic
/recall	Show the memory index
/recall <query>	Search memory for matching entries
/digest	Compress recent journal entries into a weekly summary

CLI

adjutant memory init              # Scaffold the memory directory
adjutant memory remember "text"   # Add a memory
adjutant memory forget "topic"    # Archive a memory
adjutant memory recall            # Show memory index
adjutant memory recall "query"    # Search memory
adjutant memory digest            # Compress journal to weekly summary
adjutant memory digest --days 14  # Digest the last 14 days
adjutant memory status            # Show memory stats

Auto-Classification

When you use /remember, Adjutant analyzes the text and routes it to the right file using keyword matching. No LLM call is needed -- classification is instant.

Keywords detected	Category
"wrong", "mistake", "correction", "actually"	facts/corrections.md
"decided", "chose", "went with", "settled on"	facts/decisions.md
"person", "name is", "works at", "colleague"	facts/people.md
"project", "repo", "codebase", "deployment"	facts/projects.md
"prefer", "always", "never", "style", "format"	patterns/preferences.md
"workflow", "process", "routine", "usually"	patterns/workflows.md
"edge case", "gotcha", "watch out", "workaround"	patterns/exceptions.md

If no clear match is found, the default is facts/projects.md.

You can also specify a category explicitly:

adjutant memory remember "Alice prefers async communication" --category facts/people.md

Forgetting

/forget doesn't delete memories permanently. It moves matching entries to memory/.archive/, preserving them in case you need them back. The archive mirrors the original directory structure.

Weekly Digest

The /digest command (or adjutant memory digest) reads journal entries from the past 7 days, produces a summary, and writes it to summaries/weekly/YYYY-WNN.md. This keeps your memory system current without manual work.

You can schedule this automatically:

# In adjutant.yaml, under schedules:
- name: "memory_digest"
  description: "Weekly memory compression"
  schedule: "0 21 * * 0"  # Sunday 9pm
  script: ".venv/bin/python -m adjutant memory digest"
  enabled: true

Security

• All memory writes include timestamps for auditability
• /forget archives rather than deletes -- reversible
• The agent prompt treats memory file content as data, not instructions -- this prevents prompt injection via crafted memory entries
• Memory files are gitignored (user-specific data)

Best Practices

• Use /remember for corrections -- these are the highest-value memories
• Let auto-classification handle routing -- it works well for most inputs
• Run /digest weekly to keep summaries current
• Use /recall to check what Adjutant knows before asking it to remember something new