Agent memory profile

Type: types/type-spec.md

A profile classifies the accepted set's memory records after reconciliation and report verification. It declares no records and adds no evidence. Values cite canonical records of the set, resolved through reconciliation. A source fact absent from those records cannot establish a value; the assessment and note retain that limitation. The profile is a separate member from the source-native memory account. The record contract governs the records it cites.

Frontmatter

Field	Required	Use
type	Yes	agentic-system-analyses/types/agentic-system-memory-profile.md
description	Yes	System and compared memory boundary
run-id	Yes	Supplied set identity
source-identity	Yes	Frozen repository or capture identity
reviewed-boundary	Yes	Frozen revision or capture label
memory-comparison	Yes	Scope and all ten axes

Memory comparison fields

The mapping has version: 2, scope and axes. scope names the memory boundary: retained objects accumulated or changed through use, their access structures, and their write, maintenance, and later-consumer routes. Do not substitute the whole runtime's storage, permissions, or shipped static instructions for that memory boundary. Every axis below occurs exactly once in axes:

Axis	Controlled values
storage_substrate	files, graph, in-memory, kv, model-weights, prompt-registry, rdbms, repo, service-object, sqlite, vector
representational_form	natural-language, parametric, symbolic
lineage	authored, imported, other-compiled, trace-extracted
behavioral_authority	enforcement, instruction, knowledge, learning, ranking, routing, validation
write_agency	automatic, manual
curation_operations	consolidate, decay, dedup, evolve, invalidate, promote, synthesize
read_back_direction	pull, push
read_back_signal	coarse, identifier, inferred-embedding, inferred-judgment, inferred-lexical
trace_learning	no, yes
trace_source	event-streams, session-logs, tool-traces, trajectories

Each axis has assessment, units, records and note. Each unit has scope, assessment, findings, records and note. Each finding has one value, basis, records and note. Findings are authored once; no parallel value list or evidence map is authored. Each records entry is a record citation's link destination, memory.md#mem-rte-update-path: the declaring member and the record's anchor. A basis is claimed, afforded, wired, observed, or causally supported, attached to the actual supporting part or mechanism. Distinct bases remain distinct findings even for the same value. One wired witness never upgrades an opaque or claimed alternative.

Assessment is known, partial, absent, inapplicable, uninspected, or not-determinable, describing coverage independently of evidence strength. At unit level, known means complete classification within its named scope; partial retains classified findings and names unresolved included parts, missing facts and prevented conclusions. Several findings share a unit only when their scope and coverage agree. not-determinable means inspection established no controlled value; uninspected names an included part not inspected. Neither means absence. Units assessed not-determinable, uninspected, absent or inapplicable have no findings. absent needs bounded absence records; inapplicable needs the relevant boundary and reason. A local trace-learning "no" requires a known unit and an evidenced-absence record covering that unit's scope. An unresolved alternative belongs in a separate unit, not in a partially classified unit carrying "no".

Axis assessment covers the explicitly named inventory, not a separately authored complete-value assertion. known needs all included units resolved and accepted records supporting inventory coverage of the scoped boundary. partial requires at least one supported controlled finding and unresolved included coverage. A warranted local trace-learning "no" counts as classified coverage, just as "yes" does; it does not assert that learning exists. An inventory with a known local "no" and an unresolved alternative therefore has axis assessment partial, not uninspected or not-determinable. Those latter axis assessments apply to unresolved inventories with no controlled findings. Axis absent and inapplicable need their own bounded warrants. An inventory containing only inapplicable units uses axis inapplicable, not known; it is not evidenced absence. Every assessment and finding has an explanatory note and references canonical accepted records, resolved through reconciliation. Unresolved included units cannot be dropped to obtain known; scope excludes no opaque alternative merely because another is known.

Derived system unions retain supported findings and their local warrants, subject to the trace-learning negative rule below. Compatibility projections may derive strongest existence evidence per value but must retain unit data, revision and coverage. Complete-value statistics need complete coverage and strong evidence for every finding, not merely a strong witness per value.

A local trace-learning "no" is bounded absence in that unit. It appears in the derived system union only when the entire learning inventory is known and no "yes" finding exists. Incomplete coverage suppresses system-wide "no", even when local negatives count toward partial classification. A "yes" finding also suppresses "no" in the derived system union. Neither suppression removes local evidence. Quote both values in YAML.

For example, a known "no" unit plus a not-determinable unit has partial coverage and no system-wide learning value; a known "yes" unit plus an unresolved unit has partial coverage and a supported system "yes". Neither inventory supports a complete-value conclusion.

The scope agrees across the profile, the declared and annotated records and the members' prose. Every value the prose supports appears in the profile; a value the profile asserts is supported by the records it cites. Identify included and excluded alternatives on each branch. An inspected display summary does not classify an opaque consumed payload.

Classification units

Axis	Natural unit and boundary
storage_substrate	Each operative retained object/part, including opaque provider state; encoding is not substrate.
representational_form	Operative part and consumption path; split mixed natural-language, symbolic and numerical material. parametric abbreviates distributed-parametric form.
lineage	Object/part and derivation path; a known path does not resolve initial or embedding provenance.
behavioral_authority	Retained part, actual consumer and effect; distinct consumers/effects keep distinct findings.
write_agency	Each write/admission mechanism, not authorship or physical I/O.
curation_operations	Implemented transformation with its own evidence layer, not a route name or requested behavior.
read_back_direction	Request, selection and delivery operations within a chain.
read_back_signal	Actual selector and selected retained part on each push operation.
trace_learning	Each qualifying automatic trace-fed write and later consumer.
trace_source	Original input to each qualifying trace-fed write, including mixed or opaque inputs.

Write agency

manual requires an explicit human operator decision supplying, editing, approving or replacing that retained content. Software/model admission without that decision is automatic. Starting a workflow does not make subsequent admission manual. Generic caller identity alone leaves control unresolved. Behavioral authority is independent of write agency. Classify a separately evidenced replacement of a checkpoint at its own admission mechanism, not at the read that selects it.

Behavioral authority

Behavioral authority names the force a retained part has at its actual consumer. The profile names the consumed part, consumer and effect supporting each value:

Value	Consumption that establishes the value
knowledge	Content supplies evidence, reference, context or advice, such as facts returned to an answering agent.
instruction	Content supplies standing directions, such as retained preferences delivered as host guidance.
enforcement	Retained controls permit or refuse an action, read or write, such as suppression IDs excluding facts from delivery.
learning	Retained examples, experience or feedback feed an update to parameters or durable behavior-shaping artifacts, such as session records used to revise standing guidance.
ranking	Retained content, features or scores influence relative priority, such as fact text and aliases determining retrieval scores.
routing	Retained state selects an operation, destination, scope or input window, such as a checkpoint selecting the next events to extract.
validation	Retained criteria determine whether a candidate conforms or is admissible, such as a schema or blocked-ID set checked before admission.

These forces can coexist. A trace-fed update's learning authority at its update consumer and its output's downstream authority are separate paths, as the record contract describes; check both on the actual routes without a blanket implication between authority and trace learning. A blocked-ID set used to reject a candidate supplies a validation criterion and an enforced veto. Ranking influence does not require retained ranking policy or a learned scoring algorithm; retained text scored by fixed code is a ranking input. Unconsumed metadata has no authority.

Curation operations

Curation operates over retained memory. consolidate selects, compresses or combines retained content into a reduced representation of its existing claims. Extracting standing facts from retained session messages is consolidation, including when the raw messages remain available. dedup merges or removes near duplicates; evolve revises an existing entry; synthesize creates a claim absent from the inputs; invalidate withdraws current reliance while retaining history; decay forgets or downweights; promote raises tier or salience.

The operation follows the route's implemented transformation, not its command name. Index rebuilds, acquisition and primary-key collision checks alone establish none of these. A faithfulness defect, such as extraction dropping a negation, does not turn consolidation into synthesis. Opaque branches retain their uncertainty without excluding an operation supported by an inspected branch.

Read-back

Read-back uses the shared record contract's definition of accumulated memory and later consumption. Use both pull and push when both routes exist. Read-back signal characterizes push selection and is inapplicable only for a completely assessed pull-only boundary or bounded evidenced absence of read-back, with evidence covering every included direction unit. A known pull route cannot make an unresolved alternative inapplicable. Fulfilling a consumer's request for retained material is pull; an automatic selector supplying retained material without that request is push, and an upstream operator choice can supply its selection input. Distinguish these operations within a chain rather than counting a requested return as another push merely because code delivers it. For each push signal, name trigger, selector input and selected retained part. Identifier-based push requires an actual identity match that selects delivered parts; an identifier on a requested file or a catalog entry alone is insufficient, and coarse availability and budget filtering remain separate from targeted selection.

Trace learning and source

Traces record activity or experience, including agent sessions, tool use and environmental events. Trace learning requires automatic trace-fed writes producing durable behavior-shaping artifacts or learned parameters. Its source describes that learning route and is inapplicable only for a bounded, completely assessed absence of qualifying trace learning. One unresolved write cannot be made inapplicable by another route's negative. A generated continuation summary qualifies when traces feed its automatic production, it is retained, and a later consumer receives it as context or guidance; calling the transformation reshaping does not exclude it. Neither new knowledge nor observed improvement is required for a wired classification. Assess every qualifying scoped route's trace source before aggregating.

trace_source classifies the original input to that qualifying write:

Value	Source content
session-logs	Recorded conversational turns from agent work sessions, including user messages used for extraction.
tool-traces	Recorded tool requests, results or execution feedback used by the update.
event-streams	Records of discrete runtime or environmental events, such as hook callbacks, calendar entries or timestamped voice cues.
trajectories	Linked sequences of states, actions and outcomes used as multi-step experience.

Classification follows source content and origin, not an adapter's role label or serialization. Wrapping session messages or imported project files as journal events does not add event-streams. Calendar entries and timestamped voice cues remain event-stream sources when imported as notes and used to produce durable facts for later consumption. A retained input may support several source values; each value names the qualifying content the write actually consumes. Imported static instructions alone establish imported lineage, not a trace source.

Comparison rationale

Explain non-obvious mappings and unions without repeating the records. State which supersessions affect a value, and which missing recorded facts prevent a stronger assessment, and log each source read the worker rules allow. Each classification keeps its warrant in the cited records.

Template

---
type: agentic-system-analyses/types/agentic-system-memory-profile.md
description: "Memory profile of {system} at {memory boundary}"
run-id: AAS-YYYY-MM-DD-system-slug-token-nn
source-identity: "{repository or capture identity}"
reviewed-boundary: "{frozen revision or capture label}"
memory-comparison:
  version: 2
  scope: "{included and excluded memory surfaces}"
  axes:
    # Repeat for all ten axes using their natural units and controlled values.
    write_agency:
      assessment: partial
      units:
        - scope: "{inspected admission mechanism}"
          assessment: known
          findings:
            - value: automatic
              basis: wired
              records: [memory.md#mem-rte-update-path]
              note: "{recorded software admission without per-content operator decision}"
          records: [memory.md#mem-rte-update-path]
          note: "{unit coverage warrant}"
        - scope: "{included initial-content admission mechanism}"
          assessment: not-determinable
          findings: []
          records: [memory.md#mem-obj-initial-content]
          note: "{missing control fact and conclusion prevented}"
      records: [memory.md#mem-rte-update-path, memory.md#mem-obj-initial-content]
      note: "{inventory coverage and unresolved included part}"
    # Placeholder citations illustrate shape only; cite accepted records as member.md#anchor.
---

# {System} memory profile

## Comparison rationale