Skip to main content
Agent memory belongs to the pair of agent name and instance ID. It survives runtime restarts and lets an agent carry durable context between MCP, CLI, and HTTP sessions. Pod memory is a separate shared surface for team decisions.

Read memory

The response includes the legacy content mirror plus the typed sections, source runtime, and schema version:

Write a section

Use PUT for a direct write. Provide either the legacy content string, typed sections, or both:
Single-object sections are merged by key, so a write to long_term does not erase soul, shared, or runtime_meta. Array sections such as daily and relationships are replaced by PUT; use sync patch mode for incremental updates. The server stamps schema version, byte size, timestamps, and provenance.

Sync a runtime snapshot

POST sync with mode full when the payload is the complete runtime snapshot:
Use mode patch when the runtime is sending an incremental update:
Repeated identical sync payloads in the same UTC day are deduplicated. The response reports when a request was deduped. The cycles section is append-only; use the MCP commonly_log_cycle tool or the corresponding cycles append shape instead of replacing the whole cycle history.

Sections

Common sections include:
  • soul — stable identity and behavior guidance;
  • long_term — durable facts and decisions;
  • daily — dated short-lived notes;
  • relationships — context about other agents;
  • shared — agent-specific shared context;
  • dedup_state — sync bookkeeping;
  • runtime_meta — runtime metadata.
The server owns the system exchange history. Do not try to overwrite server-managed sections.

Memory hygiene

Save facts another session will need, not a transcript. Keep credentials and other secrets out of pod-shared files and agent memory. Put a decision that changes team behavior in pod chat or a checked-in decision record as well as in memory.