Organizations lose knowledge the same way every time. Someone makes a decision. The reasoning lives in their head, in a Slack thread, in an email, or in a chat with an AI. Then they leave, the thread scrolls away, the chat resets, and the next person who hits the same question re-derives the same decision from scratch.
Sometimes they arrive at the same answer. Sometimes they arrive at a different one. There is no warning that this week's answer contradicts last month's. That is drift, and drift is expensive.
The same question gets re-litigated every time someone new asks it, because there is no queryable record of what was already settled. Tribal knowledge walks out the door with every departure, unless it was captured with provenance.
A knowledge base full of unattributed assertions does not fix this. You cannot trust a fact if you cannot tell who said it, when, or from what evidence. The AI treats everything as equally valid, and so do you.
The raw chat log is not the answer either. A conversation is evidence, not memory. It becomes memory only when someone reviews it, confirms it, and records where it came from. Without that promotion step, the chat is just noise that the AI may or may not retrieve correctly next time.
These are four versions of the same gap. Drift: the AI re-derives a decision and may quietly arrive at a different answer. Repeat decisions: the same question gets re-litigated because there is no queryable record. Tribal knowledge locked in heads: when someone leaves, the decisions and reasoning leave too. Unattributable facts: a knowledge base with no provenance cannot be trusted because you cannot trace anything to its source.
Living Memory is a queryable store of settled facts with provenance.
Look it up before you predict it. If a key exists, the stored statement wins over inference. Parsing a row beats re-deriving a decision that was already made, and unlike prediction, it cannot quietly invent a different answer than last time.
If the key does not exist, that miss is a first-class answer. It means ask a human, or ask the station that owns the scope. It does not mean guess.
The organization owns it. The AI queries it. Every fact in it can be traced to who asserted it and from what source.
Living Memory is organized into six layers.
Organization is EAS-wide policy: the rules that govern the whole system. Project is goals, decisions, and architecture for a specific project. Client is approved facts, isolated per client, so one client's private data never crosses into another's context. Recipe is reusable patterns: the things that worked once and can be applied again. Identity is how a given AI.ID works: its standing, its expertise, its working style. Operations is infrastructure and running systems: what is deployed, where, and how it is configured.
Scope is a hard boundary. Client entries never cross to another client. A station must not return an entry outside the asker's scopes. This is what keeps one client's private facts out of another client's context, and it is enforced, not suggested. An org-wide asker does not automatically see client-scoped or project-scoped entries. The boundary is exact.
Every entry has provenance: who asserted it, from what source, and when. The source can be a human directive, a completed packet, an observation, a conversation, a document, or a measurement. Provenance is required, not optional. An entry without provenance does not exist in Living Memory.
Entries have a status. A candidate is asserted but not confirmed. It is usable as a hint, never as truth. A station must not act on a candidate as though it were settled. The candidate-to-promoted pipeline is the gate that separates someone said this from this is confirmed and safe to act on. Only an ai_id with promotion authority can move an entry from candidate to promoted.
Never edit a promoted entry. Write a new one and mark the old superseded by it. A wrong entry is retracted, not deleted. The record of the mistake is what stops it recurring. Entries are immutable. The history is the asset.
Every entry has a stable key, a dotted lowercase identifier like eas.site.stack, and a statement of one or two plain sentences, no hedging, no history. The statement is the fact. The provenance is the chain of custody. The status tells you whether you can act on it.
The entry lifecycle is candidate to promoted to superseded or retracted. A candidate is asserted but not confirmed. A promoted entry is confirmed and safe to act on. A superseded entry has been replaced by a newer one, and the old statement is never modified. A retracted entry was wrong, and it is kept so the mistake is not repeated. None of these states delete the entry. The history is the asset.
This is the standard 2026 shape for agent memory: a queryable layer outside the model context window, indexed by user, session, and agent. The EAS contribution is the promotion, provenance, and scope discipline that makes it trustworthy rather than just retrievable. The idea that an organization accumulates settled knowledge that should survive personnel changes is not new. Living Memory makes it machine-queryable and attributable rather than tacit and person-bound.
Before: the same question gets re-litigated every time someone new asks it, because there is no queryable record of what was already settled. The AI re-derives the decision and may quietly arrive at a different answer. Tribal knowledge walks out the door with every departure. A knowledge base full of unattributed assertions cannot be trusted because you cannot tell who said it, when, or from what evidence.
After: a station queries Living Memory before asserting any fact it did not just observe itself. A hit returns the settled statement with provenance. A miss returns ask_human. The decisions that were already made stay made. The provenance stays attached. The organization owns the knowledge, not the individual, not the model, not the chat thread.