feat: add per-agent memory summarization and retrieval to control context growth

[XD319]

Summary

Add a lightweight per-agent memory system that summarizes completed interactions and retrieves only relevant memories for future prompts.

Problem

As the simulation runs longer, agents either:

• lose continuity because old interactions are dropped, or
• become expensive / noisy if too much history is kept in prompts

This makes long-running worlds less believable and less scalable.

Proposal

Introduce:

1. memory summarization after interactions
2. memory retrieval before future prompts

Instead of replaying full history, each agent stores short memory summaries and only a small relevant subset is injected into the prompt.

Initial scope

• store short memory summaries per agent
• persist them in server-side state
• retrieve top-N relevant memories by partner / location / recency
• inject them into the prompt

Non-goals

• vector DB
• advanced ranking
• reflection
• lifecycle / aging systems
• memory UI

Acceptance criteria

• interactions create memory summaries
• memories persist outside active prompt context
• only bounded relevant memories are injected
• agent continuity improves without unbounded context growth