Skip to content

Three-Tier Memory

Arc remembers you across sessions using three tiers of local SQLite storage.

Tiers

Tier 1 — Session Memory (RAM)

Current conversation messages. Token-budget managed — older messages are truncated to fit the context window.

Tier 2 — Episodic Memory (sqlite-vec)

Past conversation chunks stored with vector embeddings. Retrieved by semantic similarity when relevant to the current query. Uses FastEmbed (BAAI/bge-small-en-v1.5, 384-dim, fully local).

Tier 3 — Core Memory (SQLite)

High-confidence facts about you — name, preferences, projects. Always injected into the system prompt. Distilled automatically from conversations by an LLM background task.

How It Flows

  1. Context composition — before each LLM call, the ContextComposer assembles:
  2. System prompt + core facts (Tier 3, always present)
  3. Relevant episodic memories (Tier 2, retrieved by similarity to current query)

  4. Recent session messages (Tier 1, as many as fit in the token budget)

  5. After each turn — background tasks:

  6. Store the turn as episodic memory (Tier 2)
  7. Every 5 turns, distill stable facts to core memory (Tier 3)

Chat Commands

/memory              # show core facts
/memory episodic     # show recent episodic memories
/memory forget <id>  # delete a core fact

Configuration

[memory]
backend = "sqlite"
path = "~/.arc/memory"
enable_long_term = true
enable_episodic = true