datdang-dev/agent-memory

Agent Memory Layer

Overview

The Agent Memory Layer gives specialist agents (Suki, Riko, Kana) persistent memory across pipeline runs. Instead of starting cold every time, agents can reference past successes and failures — vocabulary that scored well, audit patterns, and forensic shortcuts.

Memory files are append-only and read during context loading as part of the JIT payload.

On Activation

1. Determine mode: READ (context loading) or WRITE (post-audit persistence)
2. Resolve memory path at {{run_dir}}/_pipeline/{{project_name}}/agent-memory/
3. Execute appropriate mode below

Modes

READ Mode (During Context Loading — Step 3)

| Check | Action | |-------|--------| | Memory file exists? | Append to agent's context payload | | Exceeds 2000 tokens? | Summarize: keep last 10 entries + 3-line pattern summary |

WRITE Mode (After Audit PASS — Step 6)

| Agent | What to Record | |-------|----------------| | Suki (≥90 score) | Scene vibes, effective vocabulary, sensory ratios | | Suki (revision needed) | Riko's flagged category, fix applied, bad pattern to avoid | | Riko | Score, weakest/strongest categories | | Kana | Forensic cache reuse, delta item counts |

Memory File Structure

{{run_dir}}/_pipeline/{{project_name}}/agent-memory/
├── suki-memory.md
├── riko-memory.md
└── kana-memory.md

Dependencies

• Upstream: Pipeline orchestrator (gooner-alchemist) triggers READ/WRITE at appropriate steps
• Integration: JIT Context Sharding (context loading) and State Persistence (bible-sync)