LACP memory profile

Type: agentic-system-analyses/types/agentic-system-memory-profile.md

Comparison rationale

This comparison classifies implementation-supported memory routes, not observed use or benefit. All ten axes remain partial because the accepted inventory includes uninspected offline responsibilities, unknown initial admissions and external-host alternatives. A known file or consumer does not settle a different part's derivation, admission or authority. The units preserve that separation rather than presenting a complete system value set.

Operative parts and coverage

Native entries and saved conversations are distinct stores. JSON envelopes do not turn interpreted content into symbolic knowledge, and keyed filenames do not imply a KV service. Research Markdown has graph links; the hybrid index has copied chunks and vector access structures. Neither establishes a dedicated graph/vector database. Dense representations are parametric form, not evidence of weight training. Full indexed text, not its short display preview, determines the content classification (MEM-OBJ-memory-entries, RT-OBJ-messages, MEM-OBJ-research-memory, MEM-OBJ-hybrid-index).

The narrative and epochs retain separate units. Narrative has a known file, form and startup arc reader but unresolved initial admission. Epochs have automatic selection/copying and a later CLI display; this does not establish narrative generation or startup epoch consumption. Their valid shared container needs no supersession (MEM-OBJ-agent-narrative, MEM-OBJ-epochs, MEM-RTE-epoch-synthesis).

Reconciliation makes one identity amendment: the epistemic self-model object is superseded by MEM-OBJ-self-model. This profile cites that canonical object. The distinct EPI-RTE-self-model-update remains its update-route witness; the amendment upgrades no status or authority. Other overlaps are connected stages/parts, not duplicate whole routes. In particular, query ranking and learning evaluation consume retained input even without persisting new output (MEM-RTE-hybrid-retrieval, EPI-RTE-retrieval-ranking, EPI-RTE-learning-metrics).

Admission and transformation

Automatic admission is software/model selection without a per-content human decision. Enabling an apply workflow is not approving each bullet, pattern or rename. Conversely, generic callers, role strings and manual-labelled gates do not establish human decisions. Unknown narrative initialization and separately supplied checkpoint/content replacement cannot inherit automatic agency from an unrelated derived update (MEM-RTE-keyed-memory, MEM-RTE-daily-promotion, MEM-RTE-memory-maintenance, RT-RTE-learning-promotion, RT-RTE-workflow-advance).

Curation follows the transformation. Handoff/extraction/epoch copying consolidates existing content. SMS preference and tests-pass labels can introduce non-entailed assertions, supporting synthesis only on that specific update. Normalized duplicate-line filtering is distinct from mere hash collision checking. Optional semantic helpers remain uninspected. Resolution and rollback withdraw local reliance, not every copied index or hint; file moves and truncation forget locally. Similarity-based supersedes changes lack established registry persistence, so they do not support durable semantic deduplication or universal invalidation. Title renaming and index rebuild alone establish no curation value (MEM-RTE-trace-research, EPI-RTE-memory-promotions, EPI-RTE-self-model-update, EPI-RTE-consolidation-policy, MEM-RTE-memory-maintenance).

Consumers and trace-fed updates

Authority is assessed where retained parts are consumed. Text can supply knowledge/context or standing preference instructions. The same retained text/features can separately affect ranking. Blocked research IDs validate admission and enforce exclusion. Workflow tokens constrain later transitions, not software correctness. Promotion state controls re-promotion/rollback, not an established host enforcement route. Changed source can execute later, but executable force alone does not specify which controlled effect any arbitrary edit has (MEM-BAP-memory-text, MEM-RTE-trace-research, RT-RTE-workflow-advance, RT-RTE-learning-promotion, RT-BAP-autoresearch-code).

Host recovery is push. The cwd hash actually selects the delivered handoff, establishing identifier selection; recency, significance, fixed field slices and priority budget establish separate coarse selection. SMS agent/project arguments do not establish identity filtering. Explicit native reads, CLI search, hybrid ranking, metric evaluation and status reloads are pull. Dense/lexical scoring of a requested query does not establish a push signal, nor does delivery of the requested result create a second push (MEM-RTE-session-memory, MEM-RTE-keyed-memory, MEM-RTE-hybrid-retrieval, MEM-RTE-learning-hints).

Automatic trace-derived handoffs, episodes and selected self-model guidance qualify as trace learning through durable artifacts and later context delivery; improved task capacity need not be shown. Their original inputs include conversational turns and recorded execution feedback. A Stop callback is a trigger, not evidence that the extracted assistant text originated as an event stream. Daily/history-to-research chains retain session-log and tool-trace positives while preserving mixed external inbox inputs and incompletely inspected history parsing. Raw transcript persistence, generic event capture and an opaque narrative writer do not by themselves establish qualifying trace-fed updates. Epoch display is not silently assigned the episode's host guidance consumer (MEM-RTE-session-memory, EPI-RTE-self-model-update, MEM-RTE-daily-promotion, MEM-RTE-trace-research, MEM-RTE-epoch-synthesis).

The unresolved inventory is retained on every axis rather than erased to obtain known coverage. It includes partially inspected offline mirrors, ingestion/synthesis/review and maintenance branches, external MCP/QMD paths, unknown narrative/initial-content control and opaque host checkpoint internals. Missing facts prevent complete unions and complete-value statistics, not the local positives. No new bounded absence is inferred from selective inspection (MEM-RTE-memory-maintenance, MEM-RTE-vault-access, RT-RTE-extension-load, RT-RTE-host-launch).

Source-read log

No source files were read for this profile. Classification uses only the supplied accepted records, their supplied annotations and reconciliation. No named record ambiguity required an additional frozen-source read. No new evidence, record declaration or execution result is added.