Skip to main content

AI-Agent Quickstart

Use Engram through your agent first. The CLI exists, but the best experience is: ask the agent to load memory, do the work, then propose durable memory when something useful emerges.

AI-agent chat approval

In AI-agent chat, Engram approval is conversational. The agent shows refined TYPE: ... | TEXT: ... candidates first, including Light/Balanced/Strict variants for rules. Reply yes to save the exact candidates, audit to revise them, or cancel to stop. After yes, the agent uses engram save-session --accept-all with the exact approved candidates. Direct terminal CLI saves still use A/B/C unless an accept-all command was explicitly invoked.

First message in a new session

Ask:

Use Engram for this task. Load memory for: <what we are doing>.

If slash adapters are installed:

/engram load --for-agents "<current task>"

The agent should reply with a compact count line by default, such as Engram loaded: 8 memories / 24 total related memories. With slash adapters, load --for-agents is the agent-facing route.

When an agent needs a self-contained Engram usage guide, run:

engram llm

This prints the packaged llm.txt guide and does not require engram inject.

Ask the agent:

Inject Engram memory routing for this workspace, configure it, and connect this agent.

The agent will suggest running:

engram entry

To configure memory and link AI agents in a clean web UI. Under the hood, to initialize the workspace:

engram inject

To link the same agent globally, so new workspaces can load Engram global memory without running engram inject first:

engram link --global <agent-name>

For chat-native use, ask:

Install slash support so I can use /engram directly from this agent.

Daily loop

Start:

/engram load --for-agents "current task"

During work:

/engram search "topic I might be missing"

When the agent learns one durable fact:

/engram save knowledge

When the session produced several useful rules, facts, or workflows:

/engram save-session

Short form:

/engram ss

To include recent chat history the agent can actually access:

/engram save-session --query-level 3

--query-level must be a positive integer. The agent may use up to that many recent human-agent chat sessions, including the current one, and must not invent unavailable history.

Accept-all shortcut only when you truly mean it:

/engram ss -a

-a means the human explicitly approves every agent-recommended candidate. Agents must not add it by themselves.

To mine recent accessible chats and accept all generated candidates in one request:

/engram ss -a last 50 sessions

That normalizes to engram save-session --query-level 50 --accept-all.

Import existing knowledge

For a repo that already has AGENTS.md, CLAUDE.md, Cursor rules, notes, or docs:

/engram take-control --plan
/engram take-control --all

Use --plan first when you want to see selected files, skipped files, token estimates, and likely memory types.

Global memory

Use global memory for preferences that should follow you across repos:

Set up global Engram memory at <path>, then save this preference globally:
Use pnpm for package management.

The agent may use:

engram inject --global-only --global-path <path>
engram save --scope global "Use pnpm for package management."
engram link --global <agent-name>

When inject sees configured global memory, it creates or selects a user default profile for that global root so future workspaces can reuse it.

Keep it healthy

Ask the agent at the end of meaningful work:

Check Engram health, report invalid memories, and propose anything worth saving from this session.

Useful commands:

engram upgrade
engram upgrade --plan
engram verify
engram repair
engram graph "<topic>"
engram quality-check
engram archive --reason "<why>" <id-or-file>

Next steps