MCP tools: AI memory
Last updated 10 days ago
MCP tools: AI memory
Durable facts that survive across sessions — personal (user) or team-wide (org). The same entries are visible in the client under Settings → Personal → AI settings → Memory (see guide).
search_memory
Search the memory — personal and shared org facts.
read · followerMeant for the start of a task: recall conventions, preferences and earlier decisions before doing the work.
save_memory
Store a durable, reusable fact. write · user
Only store stable, reusable facts — never transient details of a single session, never credentials.
{"name":"save_memory", "arguments":{"scope":"org","kind":"convention", "text":"Commit titles start with the item number (#216.77)."}}Team memory (scope: "org")
Every AI in the organization reads the team memory. Only unproblematic conventions, decisions and stable references belong there (standard colours, terminology, tech stack, processes). Never put there: statements about individual people, confidential customer, contract or financial details, or content from meetings, calls or transcripts — from a meeting at most an unproblematic decision. When in doubt, ask first or save privately with scope: "user".
The server enforces the kind: an org entry with a kind other than convention, decision or reference is rejected. The default kind: "context" is therefore not enough for org.
Private conversation
If the client sends the x-liza-conversation header (see context headers) and that conversation is marked private, save_memory stores nothing — regardless of scope.
Rejected calls
A rejected call is not a JSON-RPC error but a regular result with ok: false:
get_org_overview
A compact overview of shared org knowledge (conventions, standard colours, terminology, tech stack). read · follower
No parameters.
Retention
How far the memory reaches back depends on the organization's plan (ai_memory_days): the last 50 days on Free, unlimited on paid plans. Older entries are hidden when reading.
Was this helpful?
Still need help? Ask the team