memory-and-agent-systems

# Memory and Agent Systems

- Episode ID: `epi-20260409-abf1f376005e`
- Scope: `shared`
- Created: `2026-04-09T20:23:27Z`

## Question

How do Claude Code's memory and multi-agent systems work together?

## Summary

Claude Code has 5 memory systems at 3 scopes. Auto-memory extracts session notes (triggers at 10K tokens, then every 5K/3 tool calls, capped at 12K tokens). Auto-dream consolidates memory during idle time via 4-phase process (Orient, Gather Signal, Consolidate, Prune). KAIROS is the unreleased always-on daemon using tick-loop prompts, 15-second blocking budget, and append-only daily logs designed as a pair with auto-dream. Memory retrieval deliberately rejects vector search — uses LLM reasoning over filenames instead. The forked agent pattern underlies all background operations, achieving 92% prompt cache reuse. TeammateTool exposes 13 operations for full team lifecycle with file-lock task claiming. ULTRAPLAN offloads 30-minute planning to remote CCR, forming a fast/slow thinking pair with KAIROS.

## Findings

- Memory retrieval by LLM reasoning over filenames outperforms vector search for structured files

Forked agents achieve 92% cache reuse by sharing byte-identical API request prefixes
KAIROS + ULTRAPLAN form a System 1/System 2 architecture for ambient vs deep deliberation
TEAMMATE_MESSAGES_UI_CAP=50 was added after 292-agent test consumed 36.8 GB

Lessons
- Auto-dream is the garbage collector — without it, memory grows stale, contradictory, and unnavigable
- Every concrete engineering limit in the codebase has a production failure story behind it
References

none

Lessons

References