Theoretical / informal — ideas without a paper
14 systems in the theoretical / informal — ideas without a paper category of the AI Agent Infrastructure Landscape, grouped by maturity tier.
Tier 4 — early / experimental (2)
- Externalization in LLM Agents Memory + skills + protocols as coupled externalization
Traces the shift from weights-as-capability to harness-as-capability; analyzes memory, skills, and protocols as three coupled forms of externalization and examines how they interact. Memory is defined as the externalization of state acro…
- From Human Memory to AI Memory (survey) 3D taxonomy: object × form × time
Eight-quadrant classification grid across personal/system, parametric/non-parametric, and short-term/long-term axes. Bridges cognitive-science memory taxonomy to LLM architecture choices, less common than purely engineering-oriented surv…
Tier 5 — theoretical / informal (12)
- Context Engineering Discipline rename
"+1 for 'context engineering' over 'prompt engineering' … the delicate art and science of filling the context window with just the right information for the next step." The LLM is "a coworker with anterograde amnesia" — cannot consolidat…
- Context Engineering — naming event Discipline crystallisation
Endorses "context engineering" as a distinct discipline from prompt engineering — what agents do with their context window (routing, compression, tool output formatting, memory retrieval injection) is engineering, not just prompting.
- Context Expansion Law Parkinson's-Law analogue for context
"Application context tends to expand to fill the context limits supported by the model." Treats agent memory as a first-class unresolved design problem rather than a solved component; explicitly defers a memory deep-dive to a future post.
- Files Are All You Need Filesystem as memory interface
Coding agents (Claude Code, Cursor) converge on the filesystem as their primary memory abstraction: conversation histories as searchable files, skills as files, retrieval via file search rather than vector DBs. Argument: LLMs are fluent …
- Four-Type Agent Memory Taxonomy Cognitive-science → LLM mapping
"Agent = LLM + memory + planning + tools." Maps cognitive memory types to LLM machinery: sensory (learned embeddings), short-term / in-context (the context window), long-term (external vector store with fast retrieval). The first widely-…
- JEPA Six-Module World Model (informal) Memory as architecturally peer module
Six-module cognitive architecture (perception, world model, cost, memory, action, configurator). Memory is a distinct addressable module with its own update rules, separate from the world model that reads/updates it — not in-context stor…
- LLM as OS Kernel Architectural metaphor
LLM is "the kernel process of a new Operating System" orchestrating I/O, code interpreter, browser, and an embedding store for files / internal memory. Context window = RAM; embedding store = filesystem.
- LLM OS Specs Memory in hardware terms
Operationalises the kernel analogy: 256-core processor at 20 Hz (tokens/sec), RAM = 128K tokens of context, filesystem = Ada-002 embedding store. Memory capacity becomes a measurable architectural constraint.
- LLM Wiki — compiled memory Compilation over RAG
Three-layer directory: raw/ (immutable sources), wiki/ (LLM-maintained compiled pages), CLAUDE.md (schema). LLM reads sources and integrates them into the wiki — updating pages, resolving contradictions — so knowledge compounds across se…
- Memory as Metabolism Metabolic governance cycle for companion knowledge
Argues that companion AI knowledge systems must treat memory as a continuous metabolic process rather than a static store. Proposes a three-tier store (raw buffer, active wiki, cold memory) governed by five operations (TRIAGE, DECAY, CON…
- Memory Poisoning as Attack Surface Security framing
Agent persistent memory is an attack surface: an adversarial document in the environment can instruct the agent's memory tool to write malicious facts, poisoning the long-term store. The more capable the memory system, the larger the inj…
- RL-Based Continual Learning as Memory Memory-in-weights via on-policy RL
On-policy RL is naturally more resistant to catastrophic forgetting than supervised fine-tuning — its online nature biases learning toward low-distribution-shift updates. Proposal: RL post-training as a viable mechanism for agents to acc…