Memory

Context = Short-Term Memory ¹

The session (“chat”) is the agent’s context (i.e. short-term memory).

We can also edit this history (e.g. in LangChain)!

But: Context window is finite and session-bound!

Long-Term Memory

Memory system that

• persists across sessions
• can be (much) larger than context window

-> injects relevant info into context

Long-Term Memory (Level 1): MEMORY.md ¹

• Policy in AGENTS.md
• Facts/logs in MEMORY.md

# AGENTS.md (memory policy)

## Memory policy
- Load `MEMORY.md` at session start.
- Append durable facts to `memory/YYYY-MM-DD.md`.
- If user says "remember this", write it down immediately.
- Promote stable preferences/decisions into `MEMORY.md`.

## Editing rules
- Keep entries short, factual, and timestamped.
- Remove stale items when contradicted by newer evidence.

Long-Term Memory (Level 2): “Clever” MEMORY.md ¹

We can post-process MEMORY.md:

• Summarize (compress)
• Prune it
• Introduce hierarchy (INDEX.md file pointing to other MEMORY.md files)

Long-Term Memory (Level 3): ¹

Retrieval-Augmented Generation (RAG) (… and fancy systems on top of it)

• Embed documents into vector space
• Compare query embedding to document embeddings (cosine similarity)
• Inject relevant passages into context

Figure from LangChain

Figure from LangChain