What's your agent's memory architecture? (A/B/C/D/E)

[xg-gh-25]

Quick question for anyone running an AI agent/assistant more than a few sessions:

How does your agent remember things between sessions?

Curious what the actual distribution looks like in practice (not what the docs say — what you actually shipped):

A) Single file (CLAUDE.md, .cursorrules, system prompt doc)

• One big file, manually maintained
• Works until it doesn't (~month 3 it's stale or bloated)

B) RAG / vector store

• Embeddings + retrieval
• Good recall, bad precision (retrieves "related" but not "needed")

C) Structured multi-file (multiple context files with ownership/priority)

• Different files for different purposes
• Maintenance split between human and AI

D) No persistence

• Each session starts fresh
• Context via paste/attach/manual

E) Something else

• Describe below 👇

Follow-up questions (if you want to share more)

1. How many sessions before it "feels like it knows you"?
2. What's the Add Ollama support (local + cloud) with automatic model detection and OpenAI-compatible fallback #1 thing your agent forgets that it shouldn't?
3. Have you ever switched tools and lost accumulated context? How painful was it (1-10)?

---

No wrong answers. Genuinely trying to map what's actually deployed vs what's marketed.