Memory

ARC's memory system stores observations, corrections, preferences, and facts that persist across sessions. Memories are scoped, typed, and scored by relevance using exponential decay.

Searching Memory

arc memory search "deployment process"
arc memory search "react hooks" --scope persistent

Search uses deterministic keyword/scope/type/recency ranking. More recent and more frequently accessed memories score higher.

Listing Memories

arc memory list
arc memory list --scope persistent
arc memory list --scope persistent --type correction
arc memory list --type preference

Memory Properties

Scopes

Scope	Lifetime	Description
session	Cleared on exit	Ephemeral, per-session observations
persistent	Survives restarts	Durable cross-session knowledge
shared	Synced across profiles	Available to all profiles via the shared layer

Types

Type	Description	Example
observation	General observation	"This project uses pnpm, not npm"
correction	Corrected behavior	"User prefers tabs over spaces"
preference	Explicit preference	"Always use conventional commits"
fact	Factual knowledge	"API rate limit is 100 req/min"

TUI Memory Management

The TUI Memory view displays real persistent memory entries with keyboard-driven browsing and search.

Key	Action
s	Search memories by keyword
d	Delete the selected memory entry
↑ / ↓	Navigate the memory list

The view shows each entry's scope, type, content preview, and relevance score. Search results are ranked by the same deterministic scoring used by the CLI.

Web Dashboard Memory Management

The Web Dashboard Memory view provides full memory management through a browser interface:

• Add entries — create new memory entries with content, scope, and type
• Search — keyword search with the same ranking algorithm as the CLI
• Delete — remove individual entries
• Filter — narrow results by scope (session, persistent, shared) and type (observation, correction, preference, fact)

All changes are broadcast over WebSocket for real-time updates across connected clients.

TIP

Memory is fully managed from all three interfaces — CLI, TUI, and Web Dashboard. The same search ranking applies across all interfaces.

Relevance Scoring

Memories use an exponential decay scoring model. Each memory has a relevance score that decays over time, with a configurable half-life. Accessing a memory boosts its score.

arc memory stats          # See score distribution
arc memory prune --threshold 0.1

The decayScore() function computes relevance as:

score = base_score * (0.5 ^ (age / half_life)) + access_boost

Memories below the prune threshold are removed. Check the distribution with arc memory stats before pruning.

Auto-Extraction

ARC can automatically extract memories from agent conversations. The extractMemories() heuristic detects:

• Corrections — when the user corrects an agent's behavior
• Preferences — explicit user preferences mentioned in conversation
• Patterns — recurring patterns in tool usage or workflow
• Decisions — architectural or design decisions

Extracted memories are stored as persistent scope entries and surface in future sessions when relevant keywords match.

Storage

• Session memory — in-memory Map-backed, cleared on process exit
• Persistent memory — JSON file-backed at ~/.arc/memory/
• Shared memory — linked via junction/symlink to ~/.arc/shared/memory/

The web dashboard exposes memory via the /api/memory REST endpoint with full CRUD support — add, search, delete, and filter. See Web Dashboard for details.