Appendix A — Glossary

This glossary covers only terms that carry a specific meaning in this book. General engineering vocabulary (such as "interface" or "deployment") is not included. For every English term that appears in the body text, the glossary gives the unified rendering used throughout.

Term	English	Definition	Confusable with
Single-pass ADD-only	single-pass ADD-only	Writing only appends new facts; one model call completes the write and never modifies or deletes existing facts	Not the same as an append-only log: the former still performs structured extraction and typing, the latter stores raw text
Dual time axis	dual time axis	Separating "when the fact held" (observation_date) from "when the system learned it" (write time)	Not the same as a version chain: a version chain records multiple changes to one entity; the dual time axis records two temporal dimensions of one fact
Scope	scope	The isolation dimension of memory; in this book specifically the four dimensions user_id / agent_id / run_id / app_id	Not the same as namespace: scope is the logical isolation dimension, namespace is the physical carrier
Namespace	namespace	The independent storage unit derived from a scope, corresponding to an independent directory and database file	Do not mistake a "hard-coded namespace" for "isolated"
Physical isolation	physical isolation	One independent file per scope, rather than a shared table with filter conditions	Contrasted with logical isolation, whose risk is a missing filter condition
Hash-chained audit ledger	hash-chained audit ledger	Each record contains the SHA-256 hash of the previous record, forming a chain in which tampering can be located	Not the same as an operation log: a log can be deleted or altered without trace, a hash chain cannot
Provenance tracing	provenance tracing	Following the pointer chain from a high-level conclusion back to the original message, answering "which sentence did this conclusion come from"	Not the same as log tracing, which answers "what happened and when"
Pointer-chain backtracking	pointer-chain backtracking	Following the L3 → L2 → L1 pointers layer by layer to retrieve content and the covered message ids	The engine does not store L0 raw text; this is a deliberate privacy boundary
Progressive disclosure	progressive disclosure	Returning a directory first (title, one-line summary, expansion cost) and expanding the body only on demand	Not the same as pagination: the former decides expansion by semantic relevance, the latter slices by position
Budget-aware retrieval	budget-aware retrieval	Choosing which memories to return, and at what granularity, under a token budget constraint	Hard caps differ from soft caps: a hard cap discards by layer, a soft cap may truncate
Lossless compression (AIC)	lossless compression (AIC)	Compressing memory into "pointer + structured facts + content atoms", with numbers, dates, amounts, and model numbers intact at every tier, restorable byte-for-byte	Not the same as summary compression, which is irreversible
Content atom	content atom	A key fact fragment preserved intact after compression, such as a number, date, amount, model number, URL, or quoted phrase	Atoms are never trimmed at any compression tier
Cascade delete	cascade delete	Deleting a memory while handling all four layers: source record, derived conclusions, retrieval cache, and dependent summaries	Not the same as single-record deletion, which handles only the first layer
Mask	mask	Replacing the body with a placeholder while retaining the record skeleton; reversible and does not break the KV-Cache prefix	Not the same as deletion: the skeleton remains in storage and cannot substitute for deletion in compliance contexts
Notary	notary	The pre-write safety pipeline: credential detection and field-level redaction, invisible Unicode detection, HTML injection detection	Not the same as retrieval-time filtering: the former ensures contaminated content never lands
Sensitivity grading	sensitivity classification	Classifying content as normal / private / secret, which determines write admission and the recall gate	The recall gate is on by default and blocks only the secret level
Conflict resolution	conflict resolution	Judging the type of a contradiction between old and new facts and choosing an action (ADD / NOOP / UPDATE / KEEP_BOTH)	Six kinds: temporal update, preference change, scope difference, precision difference, logical contradiction, source conflict
Relevance floor	relevance floor	Candidates below this threshold do not enter the result set	Not the same as top-k truncation: the former may return fewer than k or even zero
Lexical fallback	lexical fallback	When the embedding service is unavailable, BM25 and other lexical signals provide baseline recall	Guarantees degraded availability, but recall quality is below full fusion
Sleep-time compute	sleep-time compute	An architectural idea in which a background process organizes and consolidates memory while the main flow is idle	This book makes no assertion about any system's current implementation
Seven-factor session summary	seven-factor session summary	A fixed structure for session summaries: goal, constraints, decisions, open items, artefacts, failures, preferences, with verbatim anchors	Open items are stored independently rather than mixed into the summary
Preconditions	preconditions	The mandatory must / must_not / typical_env recorded when saving a reusable experience	An experience missing preconditions causes negative transfer when conditions do not match
Negative transfer	negative transfer	An experience valid in one scenario applied indiscriminately to an inapplicable one, causing an incident	The mitigation is precondition validation at recall time
Golden dataset	golden dataset	A high-quality evaluation asset containing inputs, trajectories, results, and criteria	Not the same as a test set: a golden dataset includes trajectories and criteria, not only question-answer pairs
Trajectory	trajectory	The complete execution record of a task from goal to result, including tool calls and intermediate decisions	Not the same as a trace: a trace is an observation record, a trajectory is reusable behavioural evidence
Human-in-the-loop	Human-in-the-loop, HITL	The architectural principle that critical steps must be confirmed by a human before continuing	Not the same as an approval switch: the former is an architectural principle, the latter a feature
Retrieval-Augmented Generation	Retrieval-Augmented Generation, RAG	Retrieving external knowledge and injecting it into context to generate an answer	Not the same as memory retrieval: the former's authority comes from its source, the latter's from its process
Controlled self-evolution	controlled self-evolution	Turning effective experience into memory, experience entries, tools, and runtime mechanisms, and making them effective only after evaluation and gating	Not the same as self-modification: the former must pass reproducible evaluation and a rollback-capable release
Maturity levels	L1–L4	Four levels graded on two axes: the responsibility memory carries and how verifiable it is	Autonomy is an independent parameter orthogonal to the grading, not a grading criterion
Tier gate	tier gate	The mechanism by which the hosted edition narrows read windows according to subscription tier	Off by default; self-hosted and PyPI users are unrestricted

A.1 Two notes on terminology

First, "memory" and "context" are not synonyms in this book. Context is the input to the current inference, with a lifetime bounded by one call; memory is state persisted across calls, with a lifetime bounded by the user or task. Conflating them leads to inferences such as "a bigger context window removes the need for memory".

Second, "deletion" in this book always refers to one of the four semantics, and which one is determined by context. Where ambiguity is possible, this book states the layer explicitly (mark-delete, mask, cascade soft-delete, physical delete). Readers are advised to use the same precision when talking to vendors, because the sentence "we support deletion" carries almost no information in a compliance context.