Atlas Librarian · Memory Architecture
Seventeen memory segments, each with a distinct cognitive job.
Atlas Librarian currently defines 17 canonical memory segments. The taxonomy separates active reasoning context from stable facts, procedures, history, decisions, rules, identities, structural knowledge, relationships, time-aware follow-up and long-term retention. The result is a memory system that can reason about what a record means—not just whether its words resemble the current query.
Layer 1 · Active cognition
The memory segments closest to current reasoning and continuity.
These segments decide what should be active now, stage complete artefacts when needed, handle future relevance and preserve the handoff state required to resume work.
The front-door Librarian, index, router and segment selector. Recall should consult this layer first, then load only relevant memory into the active working set.
A deliberately small, fast workspace for the subset of memory currently participating in reasoning. It is not intended to contain bulk project history.
A separate staging area for complete artefacts such as documents, files, images or structured objects when correctness depends on reasoning over the whole object.
The temporal execution segment for due tasks, wake events, reminders, deadlines and notification-triggering follow-up. Time-aware memory is therefore distinct from passive facts.
Continuity records that transfer project or thread state across sessions, agents and restarts: milestones, execution state, current position and next steps.
Layer 2 · Knowledge and governance
What Atlas knows, how it works, what happened, what was decided and what must be obeyed.
These segments are intentionally separate because a stable fact is not the same thing as a procedure, a historical event is not automatically a current truth, and a decision record is not identical to a hard rule.
Stable project facts, truths, identifiers and durable knowledge. Semantic Memory answers questions such as what a system is, what an identifier means or which stable mapping is current.
Runbooks, workflows, repeatable sequences, repair steps and operating methods. Procedural Memory represents how work should be performed.
Historical what-happened records: incidents, completed changes, observed outcomes and session-bounded execution traces. It gives Atlas experience without turning every past outcome into a rule.
Architectural, operational and policy decisions with rationale, trade-offs, consequences and supersession history. It preserves not only what was chosen but why.
Hard constraints, non-negotiables and behaviour policies. Rules Memory carries the high-priority guardrails that should influence reasoning and action selection.
Layer 3 · Structural and relational context
Memory also needs to know where, what shape, who and how things connect.
A project is not only prose. It contains paths, contracts, named systems, people, services, dependencies and preferences. Atlas gives those dimensions their own memory roles.
Canonical paths, logical entry points, service routes and location references. Route Memory answers where something belongs or how it is reached without making that location part of an unrelated fact record.
Data shapes, contracts, fields, structural definitions and interface expectations. It preserves the shape of information separately from the business facts carried inside that shape.
Links among memory items and entities: dependencies, causes, associations and supersession chains. It supports contextual traversal beyond one isolated memory record.
Durable named identities for people, systems, services, tools, tenants and other project entities. Entity-centric memory helps disambiguate references over long-running work.
Preferences, expectations and working styles that should persist across interactions. Its default meaning is user-oriented continuity rather than universal project truth.
Layer 4 · Signals and long-term retention
Events can become history, and history can cool without disappearing.
The final segments preserve discrete signals and low-heat retained material without forcing either into the immediate working set.
Discrete events, signals, transitions and observed triggers. Event records can contribute to timeline reasoning, scheduler logic or later episodic history without being promoted automatically into permanent truths.
Low-heat retained memory that remains retrievable by explicit search, date or trace while staying out of active recall. Archive is retention without active-context pollution.
Why segmentation matters
The taxonomy is a reasoning control surface, not an exercise in naming folders.
Segment identity influences how memory is treated: whether it is eligible for fast recall, whether it can represent authoritative memory, whether it links to scheduling or artefact staging, what scope it normally belongs to and what thermal state it begins with. Those policy differences let the Librarian make memory behaviour explicit.
Facts, rules, decisions and routes often benefit from fast recall; deep historical and relationship material can remain available without always entering the first retrieval set.
A current rule or stable fact can be authoritative within the memory domain. A historical episode or user preference is contextual by nature and should not automatically acquire the same authority.
Some memory is project-wide, some belongs to one thread, some is user-specific and some is system or global context. Segment defaults provide a disciplined starting point.
A hard rule may begin Hot while event history begins Cooler and archive material is Frozen. Thermal state is explained separately in the thermal-memory architecture.
A compact mental model
Index first. Work small. Preserve meaning. Keep history retrievable.
The 17 segments can be remembered as four families: active cognition chooses and stages what matters now; knowledge and governance preserve facts, methods, history, decisions and constraints; structural context explains where things live and how they relate; event and archive memory preserve the long tail without overwhelming current reasoning.
Seventeen canonical segments