Part II · Context, Memory, Files, and Composition

Memory Ecology and Multi-Resolution Recall

Memory is not one archive that the system searches all at once. It is a set of sources, files, decisions, plans, relationships, and summaries kept at different levels of detail, so current work can recover only the history and evidence it needs.

Memory ecologyRecall ladderSemantic standingForgetting

1. Memory is an ecology

AIOS does not define memory as a database of past statements or a single vector index. Memory is a coordinated ecology of durable sources, artifacts, semantic records, projections, active branches, and retrieval paths.

Different memory forms preserve different things:

exact evidence preserves what was originally available;
canonical artifacts preserve what the person is making;
companion metadata records preserve identity, standing, structure, and lineage;
plans preserve intended outcomes, dependencies, and commitments;
decisions preserve authority, rationale, alternatives, and reopen conditions;
traces preserve the consequential movement without exposing hidden reasoning;
receipts preserve what was authorized and what happened;
syntheses preserve high-level contribution and navigation;
branches preserve unresolved alternatives and re-entry conditions;
registries preserve accepted definitions and reusable expertise;
projections make the durable field usable at a particular resolution.

No one memory form is sufficient.

Memory is therefore not a new center for the system. It is the local continuity through which human purpose, prior judgment, exact evidence, and unfinished work can become available to a later cognitive movement. The model does not own that continuity, and retrieval does not decide what is true or what may change.

2. Memory layers

Recall across memory resolutions

How memory remains available at multiple resolutions while selected durable ground returns to one bounded working context.

flowchart TB A["Attention memory\ncurrent object and movement"] W["Working context\ncomposed ground for one judgment"] L["Local movement memory\nrecent traces, open questions, returns"] D["Document memory\nartifact, Home, Map, annotations"] P["Project memory\npurpose, plan, outputs, relationships"] E["Expertise memory\nprinciples, roles, workflows, exemplars"] R["Registry and Atlas memory\nidentity, lineage, impact, evidence"] X["Exact source and archive\noriginal evidence and superseded lineage"] A --> W W --> L L --> D D --> P P --> E E --> R R --> X X -. "exact descent" .-> W E -. "selected practice" .-> W P -. "bearing" .-> W D -. "artifact state" .-> W

Memory ecology · canonical08-memory-ecology-and-multi-resolution-recall--m01.mmd

Text equivalent

Attention memory — current object and movement → Working context — composed ground for one judgment; Working context — composed ground for one judgment → Local movement memory — recent traces, open questions, returns; Local movement memory — recent traces, open questions, returns → Document memory — artifact, Home, Map, annotations; Document memory — artifact, Home, Map, annotations → Project memory — purpose, plan, outputs, relationships; Project memory — purpose, plan, outputs, relationships → Expertise memory — principles, roles, workflows, exemplars; Expertise memory — principles, roles, workflows, exemplars → Registry and Atlas memory — identity, lineage, impact, evidence; Registry and Atlas memory — identity, lineage, impact, evidence → Exact source and archive — original evidence and superseded lineage.

The vertical ladder represents changes in recall resolution, not ranked stores. The dotted paths matter: exact sources, expertise, project bearing, and artifact state are selected back into a bounded working context rather than loaded all at once.

The diagram is not a storage hierarchy. It shows possible resolutions of recall.

3. Functional memory classes

Memory class	Preserves	Typical durability	Typical use
Exact source	Original wording, evidence, data, provenance	Long-lived or archival	Verification, quotation, correction
Canonical artifact	The actual document or deliverable	Long-lived and mutable	Writing, editing, delivery
Companion metadata	Stable identity, accepted structure, standing, lineage	Long-lived and synchronized	Context, navigation, reconstruction
Working context	Selected ground for current work	Rewritable and disposable	One seat or turn
Trace	What movement occurred and what it contributed	Durable but compact	Continuity and landing
Decision record	Choice, rationale, authority, alternatives, reopen conditions	Durable	Planning, correction, audit
Plan memory	Outcome, route, dependencies, commitments, foreground	Durable and revisable	Coordination and long-horizon alignment
Relationship record	How objects bear on one another	Durable and scoped	Cross-file context and Atlas
Receipt	Authorization, input state, effect, result, failure	Durable at consequential seams	Recovery, evidence, debugging
Synthesis	High-level contribution and state	Rebuildable or accepted by contract	Navigation and resumption
Branch record	Unresolved alternative, state, re-entry condition	Durable while meaningful	Parking, exploration, future correction
Expertise resource	Reviewed reusable method or domain ground	Versioned and governed	Future context composition
Derived index or Atlas	Resolved identity, joins, impact, navigable projections	Rebuildable	Search, inspection, validation

4. Attention and memory are different

The person may move rapidly among files, questions, and projects. The system should not require all unattended work to remain in active context.

Activity follows the person's attention. Continuity does not.

When attention moves:

new input binds to the newly active object;
prior authorized work remains bound to its original object;
each file retains its purpose, state, returns, and open lines;
the current capsule is rebuilt for the new focus;
returning to a prior file restores its relevant semantic situation.

The person ordinarily acts through one immediate frame at a time. Memory lets the wider system preserve an active semantic situation for each meaningful line of work and retain the durable domain ground from which those situations can be rebuilt. These resolutions can remain available concurrently without remaining in the person's foreground.

That concurrency does not mean that one model invocation sees the whole memory ecology by default. Context composition selects a bounded projection for the movement at hand. When the immediate frame, active semantic situation, and durable domain ground are well aligned, the resulting cross-resolution context lock is hypothesized to reduce how much purpose, history, and relationship structure must be reconstructed during inference. The lock is a degree of situated alignment, not exhaustive recall; exact descent remains available when the judgment requires more ground.

Memory therefore absorbs much of the reconstruction cost of free attentional movement while leaving selection, judgment, and consequential authority visible.

5. The recall ladder

The human and model can use the same descent path:

Workspace or corpus view
  → project and relationship view
    → file synthesis and companion metadata
      → Document Map
        → movement anchor
          → exact passage, annotation, decision, or source

This supports two-way movement.

Descend for precision

Move from synthesis toward exact source when wording, evidence, or mutation matters.

Ascend for bearing

Move from exact detail toward document, project, and purpose when deciding relevance, consequence, or next work.

Good memory is not merely recall. It is the ability to change resolution without losing the relationship among levels.

6. Memory at the scale of a living file

A living file can be remembered through five aspects.

Content

What the file presently says.

Intention

Why it exists and what it is becoming.

State

Whether it is opening, developing, forming, testing, accepted, active, complete, stale, or superseded.

Relations

What it supports, depends on, contradicts, extends, replaces, informs, produces, or belongs within.

Possibility

Which meaningful movements are now available: continue, deepen, test, connect, form, decide, branch, act, integrate, or close.

These aspects let the system resume a file as a cognitive object rather than as bytes with a modification time.

7. Memory is reconstructed, not loaded wholesale

Reconstructing memory for one judgment

How a current purpose drives selective recall, judgment, and a durable update that can inform the next reconstruction.

flowchart LR O["Current object and purpose"] --> Q["Determine judgment and resolution"] Q --> I["Read maps, standing, and indexes"] I --> S["Select exact sources and memory regions"] S --> C["Compose working context"] C --> J["Perform judgment"] J --> U["Update accepted ground and canonical records"] U --> I

Reconstruction08-memory-ecology-and-multi-resolution-recall--m02.mmd

Text equivalent

Current object and purpose → Determine judgment and resolution; Determine judgment and resolution → Read maps, standing, and indexes; Read maps, standing, and indexes → Select exact sources and memory regions; Select exact sources and memory regions → Compose working context; Compose working context → Perform judgment; Perform judgment → Update accepted ground and canonical records; Update accepted ground and canonical records → Read maps, standing, and indexes.

The loop closes through accepted ground, not through preservation of the active context. Maps and standing narrow where to descend; exact regions are selected only for the judgment, then the resulting durable change refreshes what later recall can find.

The active context can disappear after the judgment because its relevant products have been integrated into durable memory.

Ephemeral coordination may be rebuilt. Canonical files and consequential records preserve the position.

Candidate discovery is plural. Depending on the operation, the system may use an exact path, lexical search, metadata filters, typed relationships, vector similarity, reranking, or direct reading. These methods solve different problems and may be combined. Each remains an access path that must preserve source identity and permit exact descent; none becomes the memory system by itself.

8. Resumption memory

When a person returns, a useful resumption account can contain:

why the work exists;
current outcome and artifact state;
the last consequential movement;
accepted decisions and active commitments;
open questions and tensions;
parked branches whose conditions now bear;
asynchronous returns that arrived while unattended;
what changed in relevant sources;
likely next movements;
the person's current input.

The goal is not to retell the project. It is to restore the person's ability to act intelligently.

9. Memory, epistemic standing, and temporal state

Memory must preserve not just content but status.

Consider three sentences:

“The source reports that conversion cost exceeded the estimate by 30 percent.”
“The prior review inferred that deep floor plates were the primary cause.”
“The project decided to exclude buildings above a specified depth.”

These represent evidence, interpretation, and decision. A memory system that stores only their semantic similarity loses the distinction that future reasoning needs.

AIOS therefore treats standing, provenance, scope, currency, and authority as part of memory. Provenance establishes lineage: where an item came from, which revision was used, which trust domain it crossed, and how it can be rechecked or recovered. It does not establish that the item is true.

Durable memory is also a temporal promotion system. It must distinguish when something was observed, when it was recorded, the period for which it is claimed to hold, and whether later ground qualified or displaced it. Its core transitions include:

observed content
  → proposed memory or relationship
    → reviewed acceptance decision and authorization to canonicalize
      → exact canonical write and post-state verification
        → settled accepted standing
          → selective activation in a later context

at any stage: correct · qualify · supersede · revoke · preserve for history

These states are not a requirement that every observation become knowledge. Most material can remain temporary, be rejected, or return to dormancy. Promotion requires appropriate evidence and authority; correction preserves the prior lineage instead of pretending that the earlier state never existed.

The system can therefore answer “What governs now?” separately from “What was accepted or observed at that time?” without deleting the history needed to understand the difference.

10. Relationship memory

Files become a knowledge field through explicit relations such as:

derived from;
used as source;
composed into;
branched from;
supports;
qualifies;
challenges;
depends on;
supersedes;
produced by;
accepted through;
returned to.

Exact handles can be structured. The intellectual account of the relation remains natural language.

Relationship memory supports retrieval by consequence, not only by similarity. A decision can retrieve the evidence it depended on; a source change can identify affected claims; a plan node can locate the artifact expected to satisfy it.

The analogy to human associative recall is useful only at this functional level: an explicit relation can serve as a cue that brings related material into consideration. A metadata edge is not a synapse, and traversing a graph is not a model of biological remembering.

11. Memory for correction

A self-correcting system needs to remember:

the prior accepted claim or decision;
the evidence and rationale that supported it;
dissent or alternatives preserved at the time;
what new evidence challenges it;
who has authority to change it;
which plans, artifacts, workflows, or commitments depend on it;
the result of review;
supersession and recovery conditions.

Without this lineage, “updating memory” risks replacing one answer with another while losing the reason for either.

The delayed write surface

Memory can carry a failure forward. An untrusted file, retrieved passage, tool output, or model-generated interpretation may appear harmless in the present turn yet become influential later if it is silently written into a relationship, summary, preference, or reusable procedure. A later context may then treat the planted item as settled ground and use it to shape a consequential judgment.

AIOS therefore separates observation, proposed memory, the acceptance decision and authorization to canonicalize, verified canonical change, settled accepted standing, and later activation. Trust domain and source identity travel with the item. Retrieved memory cannot expand permissions. Promotion is reviewable, accepted changes are reversible where possible, and derived indexes can be rebuilt after contamination without destroying the canonical sources needed for recovery.

12. Memory for delegated work

Every independent assignment should be understandable without ambient conversation.

Its memory grant includes:

parent purpose and bearing;
exact charge;
granted ground and standing;
deliberately withheld ground and reason;
scope, tools, mutation ceiling, and privacy boundary;
expected product and receiving object;
evidence, assumption, limitation, and dissent requirements;
latest useful return point;
stop condition and bounded-unknown behavior.

The return then becomes a memory object with findings, evidence, limitations, delta, and reintegration needs—not an unstructured message from another agent.

13. Memory for reusable practice

The Expertise Layer preserves approved:

principles and beliefs;
roles and domain perspectives;
workflows and stage guidance;
templates and exemplars;
terminology and metadata conventions;
accepted evidence about what worked and under which conditions.

These resources remain dormant until context composition selects them for a specific judgment.

This creates accumulated expertise without a persistent expert process or one giant always-loaded memory.

14. Purposeful forgetting and memory health

Healthy memory requires governed non-inclusion.

Candidates for release or dormancy include:

rejected context candidates;
duplicated summaries;
obsolete process notes;
expired assumptions with no continuing rationale;
failed drafts whose lesson is preserved elsewhere;
temporary scaffolds with no evidence of reuse.

Material that should remain recoverable includes:

exact sources supporting accepted claims;
material contrary evidence;
decision rationale and authority;
supersession lineage;
unresolved dissent;
the only recovery path to consequential work.

The system should be able to explain why something left routine attention.

15. Memory is not authority

The following do not make a memory item governing:

being stored;
appearing often;
being retrieved frequently;
being written confidently;
being promoted in a user interface;
having metadata;
being generated by a specialized model;
belonging to an old successful project.

Authority comes from source standing, accepted decisions, explicit grants, and governed promotion within scope.

Provenance can show who asserted a claim and how it entered the system. Frequency can show that it recurs. Neither confers truth or authority. Those require evidence, adjudication, and the appropriate human or institutional grant.

16. User experience of multi-resolution memory

The person should be able to:

understand a document before opening it;
locate work from another session;
inspect how files derive from or challenge one another;
see current decisions and unresolved tensions;
return to exact evidence when a summary is questioned;
reopen a parked branch when its condition arrives;
resume work without reconstructing the transcript;
inspect which accepted practices shaped a result;
remove a derived index and rebuild it without losing canonical artifacts, sources, or accepted standing.

The machinery can remain quiet. The user should see meaningful states, relationships, and descent paths rather than internal IDs and parser traces.

17. Research evaluation

Recall fidelity

Can the system recover exact evidence behind a compressed claim?
Does it preserve standing and revision identity?

Resumption quality

How quickly can a person continue after interruption?
Does the resumed action align with current purpose and commitments?

Correction quality

Does contradictory evidence reach every affected context?
Does superseded ground stop governing without being erased?

Cognitive load

Does the person spend less effort restating history and managing files?
Does multi-resolution navigation reduce or increase orientation burden?

Memory hygiene

Are stale and duplicate projections detected?
Can indexes and maps be rebuilt from canonical sources?
Does purposeful forgetting preserve necessary lineage?
Does untrusted or merely observed content remain outside accepted memory until promoted?
Can contaminated or superseded memory be rolled back without erasing source history?

Comparative conditions

Compare against transcript-only memory, flat retrieval, manually maintained notes, and one-summary-per-project systems.

18. Research connection: memory is an update system and a write surface

AIOS proposes temporal state, governed promotion, and source-preserving correction because durable memory must remain usable across changing work. Current research supports that design pressure. LongMemEval shows large losses in sustained-interaction memory and the importance of time-aware retrieval. MemoryAgentBench separates retrieval from learning, long-range integration, and selective forgetting, exposing pronounced weakness in multi-hop updates. Security studies then show why promotion is consequential: AgentPoison demonstrates direct poisoning of persistent agent memory, while MINJA demonstrates a query-only path by which an interaction can induce a later memory attack.

These studies support explicit time, promotion, trust, rollback, and source descent. They do not show that the AIOS ecology is sufficient, that more memory is always beneficial, or that provenance makes stored claims true. The benchmarks and attacks remain bounded representations of a wider operational problem.

Read deeper in the normalized relational-memory brief, full relational-memory memo, normalized context-composition brief, and full context-composition memo. Chapter 09 specifies the local file substrate that gives these distinctions durable form.