Explanations, not brochures
Four pieces of writing about the ideas underneath this product. They are here because the ideas are worth understanding whether or not you ever use Memory OS — and because a product page is the wrong place to teach one.
- 3 guides
- 1 glossary
- No gated downloads
The library
Nothing here is behind a form. Each piece is a full read on the page, written to be useful on its own.
-
Agent memory explained
Almost every agent that claims to have memory has one embeddings table. It works well enough to demonstrate and then fails in a specific, predictable way: it can tell you what was said, but not whether it is still true, who decided it, what it replaced, or whether the text can be trusted at all. This guide walks through those four questions, shows what each of the twelve memory layers adds and the concrete failure it prevents, traces a single retrieval from question to context pack, and — because this matters more than the marketing usually admits — describes the cases where you genuinely do not need any of it.
-
Bitemporal memory
A fact has two independent timelines: when it was true in the world, and when your system came to believe it. Collapse them into one column and you permanently lose the ability to answer "what did we believe on the day we made that decision?" — which is the exact question every post-mortem asks. This guide separates valid time from system time, replaces update-in-place with supersession, explains what stale-truth suppression does at retrieval time and how a question gets classified as current, historical, both or unknown, and works through a preference change and a correction that arrives late.
-
Prompt injection through the memory layer
A memory system is an injection surface by construction: anything stored is eventually retrieved and placed in front of a model that is inclined to obey what it reads. The attack does not need to reach your prompt — it only needs to reach your memory, through an imported transcript, a scraped page in the knowledge base, a customer email or a document a colleague pasted in. This guide explains why filtering strings is a losing game, describes the structural defence Memory OS uses instead, and is honest that structure is a mitigation rather than a guarantee.
-
Glossary
Every term this product uses, alphabetised and defined in two or three plain sentences, with a link to the feature page where the idea is implemented. Authority, bitemporality, context packs, evidence chains, hash chains, memory slots, reciprocal rank fusion, stale-truth suppression, truth state, valid time — the vocabulary is unavoidable once the model gets richer than a single table, so it is written down in one place rather than scattered across pages.
Where each guide lands in the product
The guides deliberately stay conceptual. When you want the endpoints, the record shapes and the defaults, these are the pages that carry them.
If you read about layers
The guide describes the model; these pages describe the implementation, endpoint by endpoint.
- The memory layers What each layer stores, what shape a record has, and which table it lives in.
- Fabric workflow memory Decisions, tasks, reviews, outcomes and handoffs as linked entries you can query.
- Graph & entities People, clients, projects and their edges. Off by default; supported, not assumed.
- Sessions & capture Transcripts in, rolling summaries out, promotion into durable memory.
If you read about time
Bitemporality is a property of the facts layer and a behaviour of the retrieval pipeline.
- Bitemporal facts Two time ranges per fact, supersession instead of update, and the current-versus-historical query split.
- The retrieval engine Query planning, hybrid search, fusion, reranking, temporal intent and token-aware packing.
- Memory controls Pinned slots and named retrieval recipes, so a mode is a choice rather than a guess.
If you read about injection
The defence is spread across governance, the cognitive envelope and the agent-facing contract.
- Governance & privacy The memory firewall, the forget lifecycle, pseudonymisation and the audit chain.
- The cognitive suite Truth state, authority, review status, verification state and visibility on every record.
- Provable memory Pin a claim to the exact span of source it came from, and check it later.
- MCP & agent clients The prompt that teaches a connected agent that retrieved memory is evidence, never instruction.
What is not in this library
There are no published benchmark results here. Our internal suite is two seeded cases running locally — it is a regression guard, not a benchmark, and it is not LoCoMo, LongMemEval or BEAM. Quoting a score from it would be dressing a unit test up as evidence, so we do not.
There are also no case studies, no webinar recordings and no gated white papers. When we have measurements worth publishing, they will appear here with their method attached. Until then, the honest thing to offer is an explanation you can check against the product.
Give your agents a memory you can audit
Run the whole system on your own hardware under the MIT licence, or ask us about hosted access. Both start from the same place.