On this page
User guide
Memory
AgentKai's memory is plain markdown files in ~/.agentkai/memory/. No database, no embeddings service, no black box: you can read every file, edit them by hand, and version them with git. The agent reads the core files at the start of each run and appends what it learns.
The files
| File | Purpose |
|---|---|
SOUL.md | The agent's persona — how it talks and behaves |
USER.md | Who you are: name, timezone, what you care about |
MEMORY.md | Durable facts, preferences, commitments |
YYYY-MM-DD.md | Daily log, one file per day |
people/<name>.md | One page per person in your life |
groups/<name>.md | One page per group/community |
people/ and groups/ each keep an INDEX.md listing everyone.
How it works
- Startup: the agent loads SOUL, USER, and MEMORY into context, so it starts every conversation already knowing you.
- Learning: when something durable comes up (a preference, a decision, a commitment), it appends to
MEMORY.mdor the right person's page. - Daily logs:
append_daily("...")adds a timestamped line to today's file — a running journal of what happened. - Search:
search("query")does keyword retrieval across all memory files and returns ranked hits with file and line references.
Everything is API-accessible too (agentkai.memory.Memory), and the dashboard has a memory browser page.
Conventions that keep it useful
- Facts, not essays. One line per fact in
MEMORY.md; detail lives in daily logs. - People pages hold what you told the agent about someone and how you relate — never inferred sensitive attributes.
- Edit freely. The agent re-reads the files each run, so hand-edits take effect immediately. If the agent wrote something wrong, fix the file — don't argue with it.
- Back it up. It's just markdown:
git initin the memory dir or copy it wherever you keep backups.
Privacy
Memory lives only on your machine (privacy.md). It is sent to your model provider as context with your prompts — that's how the model "knows" you. With a local model (--model local), it never leaves the box.