Memory record

Draft reference. Field names, values and rules may change before release.

Values on this page are given in English. They describe the meaning of each value; exact wire spellings are not part of this draft.

Common record fields

Every canonical record carries these fields. Names are snake_case.

Field	Meaning
id	Opaque, permanent identifier. Encodes no place, region, device or model
owner	Access reference to the person the record belongs to
domain	Life area, plus an optional project label
data_class	Data class
retention_class	One retention class; see below
provenance	Provenance label. A combined value takes the worst label
recorded_at	Server-assigned record time
version, supersedes	Version and link to the previous version
status, deleted_at	Lifecycle state and soft-delete marker
produced_by	The producer (you, One or a sub-agent instance), with a reference to the model and configuration version used

Records are never changed in place. A correction is a new version with supersedes; the previous version becomes invalid.

Relations

Relation	From	To
derived-from	Memory record; routine; goal	Episode, observation; evidence records
about	Memory record	Person reference, goal, routine, place or external object
supersedes	New version	Previous version (not retroactive)
serves	Task, suggestion, routine suggestion	Goal
concerns	Interaction	Task, memory record, routine, goal, or another product's object
references	Any One record	An Access, Work, Relay, Executor or Pay object, or an external source

No relation carries authority.

Storage shape

Table	Content
Episode	The raw input unit a record is derived from: a conversation fragment, an explicit "remember this", a reference to an event read from a connected source, an imported item. Carries source kind and provenance
Fact	A typed memory record linked to one or more episodes
Relationship	A typed, bitemporal edge between two entities ("manager of, since a given month"), linked to one or more episodes

A record whose source cannot be shown is not written. Every vector carries the name and version of the embedding model that produced it; vectors from two model versions are never mixed in one search.

MemoryRecord fields

Field	Values	Notes
Content type	preference, routine, goal, constraint, relationship, personal fact, opinion, event	Extensible; each new type maps to a core type
Epistemic state	said, observed, inferred, hypothesis, verified, invalid	How One knows it
Layer	event, consolidated fact, behaviour rule
Time kind	permanent, for now, seasonal, one-off, in the past
Valid time	valid_from, valid_to	When the fact was true in the world
Record time	recorded_at, superseded_at	When One learned it or superseded it
Source episodes	Episode references	Cannot be empty
Source kind	user statement, user behaviour, connected source, external content, import	Feeds confidence and provenance
Evidence	Observation count, ratio, first and last observation, conflicting records	Confidence is computed from this
Alternative explanations	Typed list of candidates	When there is more than one reading
Life area and project	One or more areas, optional project	Retrieval filter
Sensitivity class	normal, or a sensitive class	Sets retention class and gates
Lifecycle state	candidate, active, not in use, permanently forgotten
Usage restriction	Empty, or "only for this area or project"
Version	supersedes reference	Correction chain

Epistemic states

State	Meaning
said	You said it
observed	Seen directly in your behaviour
inferred	Derived from observations
hypothesis	Untested inference with more than one explanation
verified	Confirmed by you, or by independent evidence from a different source
invalid	Replaced, or corrected by you

Imported and externally derived records arrive as candidates; their epistemic state reflects how One knows them (usually hypothesis), not what the source claimed. Two pieces of evidence from the same source count as one.

Layers

Layer	Meaning	Write rule
event	Something that happened at a point in time	Written automatically in the background
consolidated fact	Lasting knowledge merged from events	Written automatically for normal classes
behaviour rule	How One should behave	Only proposal, then your acceptance, then a new version

Time kinds

Kind	Behaviour
permanent	No estimated end
for now	An estimated end; at the end, a one-line confirmation question
seasonal	Retrieved according to its season
one-off	Does not change the lasting preference
in the past	Not used as a fact for present-tense questions

Confidence classes

Confidence is computed from evidence with a versioned formula pinned by test vectors. When the formula changes, existing confidences are recomputed and the change is recorded. Users see a class, not a number.

Class shown	Condition
"You told me"	said, or verified by you
"I think so"	observed or inferred, above the threshold
"I'm not sure"	hypothesis, or below the threshold
"I saw it in a source, not verified"	candidate from external content
"Imported, not verified"	candidate from an import

Write paths

Path	Result
Explicit "remember this"	active, said
Background inference, normal class	active; listed in "recently learned"
Background inference, sensitive class or behaviour rule	candidate until you confirm
External content	candidate; lasting only with your confirmation (labelled user.memory) or independent second evidence (labelled one.inferred); never a behaviour rule
Import	candidate, "imported, not verified"; never enters personal training
Behaviour rule proposed by One	candidate; active once you accept

Life areas

work, health, finance, relationships, home, travel, shopping, learning, leisure, general.

The list is fixed. You add projects under one or more areas, not new areas. A closed project is archived and excluded from retrieval until reopened.

Retrieval pipeline

Step	Rule
1. Task area	Limit to the task's area and project, plus general
2. State filter	Drop candidate, not in use, permanently forgotten and invalid records; apply usage restrictions
3. Policy gate	Sensitive-class consent and sharing preferences, evaluated with the shared restriction language
4. Opinion filter	Leave out opinion records for factual questions
5. Search	Hybrid lexical, vector and relationship traversal
6. Ranking	Relevance, decay priority, confidence, time kind
7. Smallest subset	Fewest records the task needs
8. Presentation	Marked data block; each line labelled with content type, epistemic state, confidence class and conflict flag

Only the output of step 7 reaches a cloud language model.

Forgetting

Verb	Effect	Reversible
Correct	New version; the old one becomes invalid	Yes
Don't use	Record kept, state "not in use"; excluded from retrieval and training	Yes
Forget permanently	The record key and every wrapped copy are destroyed; the content and its derivatives become unreadable	No

Forgetting permanently covers the record content, episode content belonging only to it, derived consolidated facts and summaries, text views, embeddings and device caches. Each record and its derivatives are encrypted with their own record key, wrapped by the person and retention-class data key; other records are unaffected. The skeleton (identifier, state, time of destruction) remains; content, content type and source do not. A personal adapter that saw the record is retrained from the remaining memory, the old version's key is destroyed, and the old adapter is not used meanwhile.

Single-record forgetting needs a clear in-app confirmation. Forgetting a life area or all memory happens on the Access approval surface with strong authentication and is never started by One.

Retention classes

Class	One data
one_memory	Memory records, episodes, behaviour rules, personality profile, life area and project definitions, tasks, attention decision records
one_sensitive	Records in opened sensitive classes, under a separate data key; made unreadable as soon as consent is withdrawn
one_observation	Raw audio, screen, camera and location; not kept by default
one_model	Personal adapter and personal small models
one_provider_trace	Which processor, region and sub-processor received which data class; metadata only
one_browser_profile	Executor browser profile, cookies and snapshots; made unreadable when the bound Mandate or task ends
one_training_lineage	Content-free lineage records for opt-in shared training

Embeddings, summaries, text views and device caches take the class of their source; a derivative of several sources takes the most restrictive one.

Export package

Part	Content
Memory records	All fields, including version chains
Episodes	The raw inputs records came from; references for data kept at the source
Personality profile and rules	Name, form of address, style, values, accepted behaviour rules and their versions
Grant summary	A readable summary of the Grants and Mandates given to One and its sub-agents, read from Access at export time
Connections	Connected accounts and capabilities, without secrets or tokens
Task history	One's single-person tasks, their steps and results
Personal adapter	Weights, base model name and version, training date, with a note that it will not work on another base model
Text view	A human-readable form of the whole package

• The schema is open, versioned and defined in CDDL; the package is encoded as JSON. Schema versions are backward compatible, and readers skip fields they do not know.
• Starting an export requires strong authentication. The package is delivered to your device, encrypted with your key. Sensitive classes are in a separate section you can leave out.
• If Access is unreachable, the export job waits; an incomplete package is never produced.

Import

• Sources: a One export package, other assistants' memory dumps (text or file), and direct transfer protocols through adapters.
• Every imported record arrives as a candidate, "imported, not verified". Its content is data; command sentences inside it are not executed and cannot become behaviour rules.
• Imported records never enter personal training, even after you confirm them.
• You can confirm imported records in bulk per area; behaviour rules and sensitive classes are confirmed one by one.