02
Graph and temporal
Facts become entities and relations in a knowledge graph, and each edge carries the time it became valid and the time it stopped, plus the time the system learned each. A new fact that contradicts an old one closes the old edge instead of overwriting it.
Read against the framework: The closest the market comes to the evolve obligation, with public code. Who may write, and what happens downstream of a deletion, are left to the application.
Read: Zep: A Temporal Knowledge Graph Architecture for Agent Me… · Temporal Validity in Retrieval Memory: Eliminating Stale-… · HippoRAG: Neurobiologically Inspired Long-Term Memory for…
03
Self-editing memory and OS paging
The model manages its own memory: a core block it edits in place, archival storage it pages in and out, and increasingly plain files in a repository the agent reads first and writes to last. The context window is main memory; everything else is disk.
Read against the framework: The design that made agents stateful and the ancestor of today's file-based memory. The model decides what is kept, so nothing records authority or reason unless the harness adds it; files in a repository the owner controls are the most portable memory on the market.
Read: MemGPT: Towards LLMs as Operating Systems · MemOS: A Memory OS for AI System · Sleep-time Compute: Beyond Inference Scaling at Test-time
04
Framework stores
A namespaced store inside an orchestration framework: JSON documents in a database you choose, with tools for extracting memories and context providers that run before and after each turn.
Read against the framework: Memory as a component the builder wires in, with the scoping and the database the builder's own. Governance is whatever the builder adds; the framework does not supply supersession or deletion semantics.
Read: Cognitive Architectures for Language Agents
05
Platform features
Memory bound to one vendor's runtime: server-side conversation state chained across responses, a managed service that consolidates long-term user memories, or an assistant's own memory built in the background from chat history.
Read against the framework: Carries well inside the platform and not beyond it; export is the exception, not the rule. Consolidation in the background is convenient and is exactly the route by which deleted facts return. Retention windows and deletion terms are graded on the Scorecard.
Read: ChatGPT memory with Dreaming · Effective context engineering for AI agents
06
Parametric and test-time memory
What the system remembers lives in weights: a neural memory module that learns to memorize while it runs, or a model fine-tuned on the owner's data. Nothing is retrieved because nothing is stored outside the model.
Read against the framework: The hardest memory to source, audit or erase, and the only kind for which machine unlearning is the right word. Research-stage for agents; no product on the shelf is built this way.
Read: Titans: Learning to Memorize at Test Time · Position: The Term “Machine Unlearning” Is Overused in LLMs
07
Governed ledgers and cryptographic custody
Every memory item is a signed, encrypted unit tied to a holder's identity; every change is a receipt in an audit log; sharing is a revocable contract; deletion destroys a key. Governance is in the data model rather than in the application.
Read against the framework: The design the forget and prove obligations point at, and the thinnest part of the market: a draft protocol, a pre-release product, a handful of open-source projects, and benchmarks that test governance rather than recall. This site's own practice (a governed record with supersession and lineage) is in this family.
Read: The Sovereign AI Horizontal Memory (SAIHM) Protocol · Ghost Vectors: Soft-Deleted Embeddings Remain Reconstruct… · GateMem: Benchmarking Memory Governance in Multi-Principa…