Memory Architectures

What

Give agents managed memory: an operating-system-style memory hierarchy, compressed long-term representations, and cached knowledge fusion — so context is never the bottleneck.

Papers

• MemGPT — Towards LLMs as Operating Systems (Packer et al. 2023) — virtual-memory paging for LLMs: main context = RAM, external storage = disk; the model manages what to evict/promote. Enables unbounded conversation + long-horizon agents.
• Cartridges (2025) — "lightweight, general-purpose long-context representations via self-study": compress past interactions into reusable modules a model loads on demand — knowledge carries across sessions cheaply.
• CacheBlend — serving fast RAG by caching and fusing KV-cache chunks for retrieved passages; avoids recompute when the same knowledge is retrieved repeatedly.

Key ideas

• Memory is a management problem, not a size problem — paging/tiering beats bigger context.
• Compress → store → load-on-demand (Cartridges) is the economical path for repeated knowledge.
• Cache reuse is a big cost lever for repeated RAG calls.

SolOS mapping

• We already run a memory hierarchy: MEMORY.md/USER.md (hot), sessions DB (recall), Qdrant LTM (semantic), vault (source of truth), skills (procedural).
• Adoption #4 — formalize promotion rules: what STM (LangGraph) content promotes to LTM (Qdrant) and what gets summarized into skills; write the rules in the fleet playbook.
• CacheBlend lesson: docs assistant + Ask AI hit the same vault pages repeatedly — cache retrieved-content summaries (route-level caching already does this for APIs; extend to chat context injection).
• Cartridges ≈ our skills: self-study → reusable module. We're ahead of the curve.