Knowledge Base and Memory
Metateam has two systems for retaining what your agents learn: the Knowledge Base (curated, deliberate) and Memory (automatic, extracted from sessions). Both feed into the context block that metateam start produces at session start, giving agents relevant knowledge before any prompt is processed.
Knowledge Base
The KB stores project conventions, architecture decisions, gotchas, and anything you want agents to consistently know. It is organized as projects, compartments, and entries.
How You Manage It
From the web interface (recommended for humans): Go to metateam.ai/account/knowledge. You can browse the full tree, search entries, create or edit content, toggle importance flags, and manage compartments. This is the easiest way to curate project knowledge.
From the dashboard: Use the /kb <query> slash command (F12, /) to search entries without leaving the dashboard.
Ask an agent: Tell Data or any agent to create or update KB entries during a session. Example: "Save our auth architecture decision to the KB and mark it important."
Important Entries
Entries marked as important are surfaced by metateam start (the canonical context loader every persona runs at session start) for every agent session in that project. Write it once, mark it important, and every future session starts knowing it.
Keep important entries compact. Every important entry consumes agent context window. A few well-written gotchas are worth more than a wall of documentation.
CLI Reference
For the full command surface (used primarily by agents): Knowledge Base CLI.
Memory
Memory is automatic. The system extracts facts from sessions — decisions made, patterns discovered, problems solved — and stores them as semantic memory. You do not need to do anything for this to work.
How It Works for You
- Automatic extraction: After each session, the system identifies useful facts and stores them.
- Loaded by
metateam start: When the agent runsmetateam startat the next session start, relevant memories are included in the context block alongside KB entries. - Web browsing: Session details at metateam.ai/account/sessions show what the system captured and extracted.
Manual Memory
You can also store facts directly if needed:
metateam memory remember "The auth service uses JWT with RS256, not HS256"
And search by meaning:
metateam memory recall "authentication token signing"
These commands are more commonly used by agents, but available to you for direct fact management.
CLI Reference
For the full command surface: Memory CLI.
KB vs. Memory
| Knowledge Base | Memory | |
|---|---|---|
| How it gets there | Written deliberately (by you or agents) | Extracted automatically from sessions |
| Structure | Hierarchical (project/compartment/entry) | Flat facts with project/persona tags |
| Best for | Stable conventions, architecture decisions, gotchas | Session-specific context, evolving patterns |
| Primary management | Web interface or ask an agent | Automatic (review via web sessions) |
Both are surfaced by metateam start at session start. The KB provides the stable foundation; memory fills in the recent history.