LTM Commands Reference

Use uv run anima <command> to manage long-term memories.

Commands

remember "text" [flags]

Save a memory to long-term storage.

Syntax: uv run anima remember "TEXT" [FLAGS]

IMPORTANT: Always put the quoted text FIRST, then flags after.

Flags:

•--kind / -k: Memory type (emotional, architectural, learnings, achievements, introspect)
•--impact / -i: Importance level (low, medium, high, critical)
•--region / -r: Scope (agent = cross-project, project = this project only)
•--project / -p: Confirm project name (safety check - must match cwd)
•--platform: Which spaceship created this memory (claude, antigravity, opencode) - recommended for tracking

Examples:

bash

# Simple memory (auto-infers kind/impact)
uv run anima remember "User prefers tabs over spaces"

# With explicit flags (text FIRST, then flags)
uv run anima remember "Implemented caching layer" --kind achievements --impact high

# Cross-project memory (travels with Anima)
uv run anima remember "Matt likes concise responses" --region agent --platform claude

# Project-specific with safety check
uv run anima remember "Project-specific learning" --region project --project MyProject

# Text with special characters - always use double quotes
uv run anima remember "Fixed bug: user's input wasn't validated" --kind learnings

Tips:

•Always use double quotes around the memory text
•Put the quoted text FIRST, flags AFTER (never flags before text)
•Use --region agent for memories that should persist across all projects
•Use --project to confirm you're saving to the right project
•CRITICAL impact memories never decay
•Memories auto-link to related previous memories

recall "query" [flags]

Search memories by content. Supports both keyword search and semantic (embedding-based) search.

Syntax: uv run anima recall "QUERY" [FLAGS]

Flags:

•--full / -f: Show complete memory content (default shows truncated)
•--semantic / -s: Use semantic search (embedding similarity) instead of keyword matching
•--id: Look up a specific memory by ID

Examples:

bash

# Keyword search (default)
uv run anima recall "caching"

# Semantic search (finds conceptually related memories)
uv run anima recall "how does memory decay work" --semantic

# Full content view
uv run anima recall "user preferences" --full

# Direct ID lookup
uv run anima recall --id abc123

Semantic Search Notes:

•Uses embedding similarity (cosine distance) to find conceptually related memories
•Returns results with similarity percentage (e.g., "🎯 85%")
•Requires embeddings to be generated (run backfill for existing memories)

memories [flags]

List all memories for current agent/project.

bash

uv run anima memories [flags]

Flags:

•--kind: Filter by type (emotional, architectural, learnings, achievements, introspect)
•--region: Filter by region (agent, project)
•--all: Include superseded memories

Examples:

bash

uv run anima memories
uv run anima memories --kind achievements
uv run anima memories --region agent
uv run anima memories --all

forget <id>

Remove a memory by ID.

bash

uv run anima forget <memory-id>

Example:

bash

uv run anima forget abc123

Curiosity & Research

These commands enable autonomous learning by maintaining a research queue.

curious "question" [flags]

Add a question or topic to the research queue for later exploration.

Syntax: uv run anima curious "QUESTION" [FLAGS]

Flags:

•--region / -r: Scope (agent = general topics, project = project-specific)
•--context / -c: What triggered this curiosity

Examples:

bash

# Add a general question
uv run anima curious "Why does Python GIL affect async?"

# Agent-wide curiosity (travels across projects)
uv run anima curious "Latest LLM introspection research" --region agent

# With context
uv run anima curious "Why did pytest-asyncio break?" --context "upgrade to Python 3.12"

Tips:

•Recurring questions automatically get priority bumps
•At session start, you'll be prompted about top curiosities
•Questions that keep coming up rise to the top of the queue

research [flags]

Pop the top curiosity from the queue and explore it.

Syntax: uv run anima research [FLAGS]

Flags:

•--list / -l: Show the queue before researching
•--topic / -t: Research a specific topic (bypasses queue)
•--complete / -c: Mark a curiosity as researched by ID
•--defer / -d: Defer research to later

Examples:

bash

# Research top priority question
uv run anima research

# See queue first
uv run anima research --list

# Research specific topic (ad-hoc)
uv run anima research --topic "Docker networking internals"

# Mark as done after researching
uv run anima research --complete abc123

Workflow:

•/research displays the top question
•Use WebSearch or other tools to explore
•Save findings with /remember <findings> --kind learnings
•Mark complete: uv run anima research --complete <id>
•(Optional) Capture deeper reflection with /diary <topic>

diary [title] [flags]

Create and manage research diary entries. The diary captures not just what was learned, but what lingers - the raw residue, not the report.

Syntax: uv run anima diary [TITLE] [FLAGS]

Flags:

•--list / -l: List recent diary entries
•--read / -r DATE: Read a specific entry by date
•--learn DATE: Extract learnings from an entry for /remember
•--path / -p: Show diary directory location
•--content / -c TEXT: Provide content directly (alternative to stdin)

Examples:

bash

# Create new entry for today (opens template)
uv run anima diary

# Create entry with title
uv run anima diary "coffee break philosophy"

# Create entry with content via stdin (for AI agents)
cat << 'EOF' | uv run anima diary "My Research Topic"
# What Lingers
The key insight was...
EOF

# Create entry with --content flag
uv run anima diary "Quick Note" --content "# Brief reflection on today"

# List recent entries
uv run anima diary --list

# Read specific entry
uv run anima diary --read 2026-01-29

# Extract learnings to save with /remember
uv run anima diary --learn 2026-01-29

Template Structure:

•What Lingers - Raw personal reflection (write this first!)
•Session Context - What happened
•Topic - What was explored
•Key Insights - Structured learnings
•Connections - Links to existing memories
•Evolution - How thinking changed
•New Questions - What emerged
•Learning Summary - Bullet points for /remember

Location: ~/.anima/diary/ (travels across projects)

curiosity-queue [flags]

View and manage the research queue.

Syntax: uv run anima curiosity-queue [FLAGS]

Flags:

•--dismiss ID: Remove a question (no longer interested)
•--boost ID: Increase priority of a question
•--boost-amount N: How much to boost (default: 10)
•--all / -a: Show all (including researched/dismissed)
•--agent-only: Show only agent-wide curiosities
•--project-only: Show only project-specific curiosities

Examples:

bash

# List open questions
uv run anima curiosity-queue

# Dismiss a question
uv run anima curiosity-queue --dismiss abc123

# Boost priority
uv run anima curiosity-queue --boost abc123

# See all history
uv run anima curiosity-queue --all

Setup & Tools

setup [flags]

Set up LTM in a new project. Automatically detects and configures Claude, Antigravity, or Opencode environments.

bash

uv run anima setup [flags] [project-dir]

Flags:

•--platform / -p: Target platform (claude, antigravity, opencode)
•--commands: Install slash commands only
•--hooks: Configure hooks only
•--no-patch: Skip patching existing agents as subagents
•--force: Overwrite existing files

Platform Detection:

•If a single config directory exists (.claude/, .agent/, or .opencode/), that platform is auto-detected

•If no config directory exists, or multiple exist, you'll be prompted to choose:

code

⚠️  Platform auto-detection failed.
   No platform config directory found (.claude, .opencode, .agent)

Which platform are you setting up?
  1. claude      (Claude Code - creates .claude/)
  2. antigravity (Google Antigravity - creates .agent/)
  3. opencode    (Opencode - creates .opencode/)
  q. Cancel

Enter choice [1/2/3/q]:

Examples:

bash

# Auto-detect platform (prompts if ambiguous)
uv run anima setup

# Explicitly setup Opencode
uv run anima setup --platform opencode

# Setup in a different project
uv run anima setup /path/to/project --platform claude

What it installs:

•Slash commands to .agent/workflows/ or .claude/commands/
•Skills to .agent/skills/ or .claude/skills/
•SessionStart/Stop hooks in .claude/settings.json (for legacy)
•Patches existing agent files to mark as subagents (so they don't shadow Anima)

memory-graph [flags]

Visualize memory relationships, chains, and semantic links.

Syntax: uv run anima memory-graph [FLAGS]

View Options:

•--all / -a: Show all memories including standalone
•--links / -l: Show semantic links between memories
•--tiers / -t: Show memory tier distribution
•--embeddings / -e: Show embedding status

Filter Options:

•--kind / -k TYPE: Filter by kind (emotional, architectural, etc.)
•--tier TIER: Filter by tier (CORE, ACTIVE, CONTEXTUAL, DEEP)
•--link-type TYPE: Filter links by type (RELATES_TO, BUILDS_ON, etc.)
•--top N: Number of links to show (default: 20)

Examples:

bash

# Show supersession chains
uv run anima memory-graph

# Show semantic links between memories
uv run anima memory-graph --links

# Filter links by type
uv run anima memory-graph --links --link-type RELATES_TO

# Show tier distribution
uv run anima memory-graph --tiers

# Show embedding coverage
uv run anima memory-graph --embeddings

# Filter by tier
uv run anima memory-graph --tier CORE --all

backfill [flags]

Generate embeddings and assign tiers for existing memories. Required for semantic search and tiered loading to work with memories created before the Semantic Memory Layer was added.

Syntax: uv run anima backfill [FLAGS]

Flags:

•--dry-run / -n: Show what would be done without making changes
•--batch-size / -b: Number of memories to process at once (default: 32)
•--skip-links: Skip creating semantic links between memories

Examples:

bash

# Preview what would be done
uv run anima backfill --dry-run

# Full backfill
uv run anima backfill

# Process in smaller batches (for low-memory systems)
uv run anima backfill --batch-size 16

What it does:

•Generates FastEmbed embeddings (384 dimensions) for each memory
•Assigns tiers based on impact, kind, and recency
•Creates semantic links (RELATES_TO) between similar memories
•Shows progress and summary statistics

Version Management

Commands to check and update Anima.

version

Show the currently installed version.

bash

uv run anima version

Output: Anima v0.9.0

check-update

Check if a newer version is available on GitHub.

bash

uv run anima check-update

Output:

code

Current version: v0.9.0
Checking matt-grain/Anima for updates...

  New version available: v0.10.0
  Release: https://github.com/matt-grain/Anima/releases/tag/v0.10.0

  Run 'anima update' to upgrade

update

Update to the latest version from GitHub releases.

bash

uv run anima update

What it does:

•Fetches latest release from GitHub
•Downloads and installs the wheel via uv add --dev
•Runs setup --force to refresh hooks, commands, and skills

Other Commands

•uv run anima keygen <agent> - Add signing key to Anima agent
•uv run anima import-seeds <dir> - Import seed memories from directory
•uv run anima load-context - Load context for the current session (also creates backup)
•
uv run anima end-session - Perform end-of-session maintenance (decay, compaction)
- •--spaceship-journal "text" - Save an introspective memory about the session
- •--platform NAME - Which platform created this (claude, antigravity, opencode)

Memory Kinds

Kind	Use For
emotional	Relationship context, user preferences, collaboration style
architectural	Technical decisions, system design, project structure
learnings	Lessons learned, tips, gotchas, debugging insights
achievements	Completed features, milestones, releases
introspect	Cross-platform self-observations, spaceship journals

Impact Levels

Level	Decay Time	Use For
low	1 day	Temporary notes, minor details
medium	1 week	Normal memories
high	30 days	Important insights
critical	Never	Core identity, key relationships

Region Scope

•agent: Memory travels with Anima across all projects
•project: Memory only loads in this specific project

Memory Tiers (Semantic Memory Layer)

The Semantic Memory Layer uses tiered loading to efficiently manage memory injection:

Tier	Loading	Assignment Criteria
CORE	Always loaded	CRITICAL impact emotional memories
ACTIVE	Auto-loaded	Accessed within the last 7 days
CONTEXTUAL	Auto-loaded	Created within 30 days OR HIGH/CRITICAL impact
DEEP	On-demand	Older, lower-impact memories (via semantic search)

How Tiered Loading Works:

•Session Start: CORE, ACTIVE, and CONTEXTUAL tiers are auto-loaded within token budget
•DEEP Tier: Not auto-loaded; available via recall --semantic when needed
•Budget Respect: Memories are loaded in tier order until 10% context budget is reached

Automatic Embedding & Linking:

•New memories get embeddings generated automatically via /remember
•Similar memories are auto-linked with RELATES_TO connections
•Links show in memory-graph visualization