learn

What is agent memory? Capture and recall, explained

Agent memory is the loop that lets an AI agent remember across sessions — capture what happened, recall the relevant slice later. Here is how dig implements it locally and reversibly.

An agent with no memory starts every session from zero. Agent memory is the loop that fixes that: capture what happened, then recall the relevant slice when it matters again — without overflowing the context window.

The two halves

dig implements agent memory as two operations over a local knowledge base:

  • retain captures content — a note, a document, or a whole finished agent session — to a dated memory/ path. It is indexed and reversible. A Claude Code transcript can be rendered to readable Markdown and retained automatically.
  • recall returns a token-budgeted, provenance-tagged context pack. Its snippets land on the passage relevant to your query, so the agent loads what it knows about a topic without blowing its window.
# capture a finished session, then recall it later
dig retain --transcript session.jsonl
dig recall "what did we decide about auth" --budget 2000

Why local and reversible matters

  • Local-first. With no model endpoint configured, dig makes zero network calls. Your memory never leaves your machine.
  • Reversible. Every capture is a journaled changeset. Memory is data you can audit and dig undo — not opaque agent state.
  • Provenance. Recalled snippets carry where they came from, so the agent (and you) can trust the source.

How agents reach it

retain and recall are exposed as dig_retain and dig_recall over MCP, and as tools in the Vercel AI SDK adapter. Any harness — Claude Code, Codex, Cursor — reads and writes the same memory.

See it in a workflow · Install dig