Internals: Auto-Memory, Auto-Dream, and Agent Teams

Source ID: src-20260409-a14e9e98c3cd
Kind: analysis
Scope: shared
Origin: community-analysis
Raw path: sources/raw/internals-auto-memory-auto-dream-and-agent-teams__src-20260409-a14e9e98c3cd.md
Status: active

Summary

Community analysis document covering Claude Code architecture, internals, and leak analysis. See extracted entities and claims below.

KAIROS (service): Unreleased always-on autonomous agent mode. Injects messages when the que
Forked Agent Pattern (concept): Shared primitive underlying all background operations in Claude Code. Creates ch
Auto-Memory (service): Background extraction system where Claude reads its own conversation and writes
Auto-Dream (service): Background memory consolidation process that runs during user idle time. Reads a
Permission Pipeline (service): Four-layer auto-approval decision stack: (1) existing permission rules, (2) acce
TeammateTool (service): Multi-agent coordination layer exposing 13 operations: team setup (spawn, discov
Tool Result Storage (concept): Mechanism for offloading large tool outputs to temporary disk files. Conversatio
Model Codenames (concept): Internal codenames for model versions (Capybara, Fennec, Numbat, Mythos, Tengu).
Agent Teams (service): Multi-instance coordination layer with 3 execution models, 13-operation Teammate
SleepTool (tool): KAIROS-only tool that lets the autonomous agent yield control explicitly, choosi
Retry System (service): 10 retries with exponential backoff, auto OAuth refresh, 90-second watchdog, and
Worktree Isolation (service): Git worktree-based agent isolation. Each spawned agent gets its own working dire

[fact] Auto-memory extraction cadence: 10K initial, then 5K/3 tools
[fact] KAIROS injects messages to keep agent alive
[fact] TeammateTool exposes 13 operations for full team lifecycle
[fact] Forked agents achieve 92% prompt cache reuse
[fact] Auto-dream runs 4 phases: Orient, Gather, Consolidate, Prune
[decision] Memory retrieval uses LLM reasoning over filenames, not vector search
[fact] Permission pipeline has 4 decision layers with ML classifier last
[lesson] TEAMMATE_MESSAGES_UI_CAP=50 added after 292-agent run consumed 36.8GB
[fact] Workers communicate via mailboxes, XML notifications, and shared scratchpad