Autogenesis memory profile

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

Comparison rationale

The profile covers retained content and its access structures, not every file or permission in the enclosing runtime. The main supported learning route is MEM-RTE-history-compaction: projected lifecycle records become model-written summaries, and later calls consume those summaries. Its wired trace-learning value does not claim improved capacity, faithful summarization, exercised reliance or causal benefit. The raw-history update consumer has learning authority; the summary's later foreground consumer has knowledge authority. These are separate paths rather than an implication from storage to learning.

The summaries' file representation does not repair the resident-state horizon. MEM-ABS-tiered-reload establishes a bounded reload absence, not a trace-learning negative. Todo reload, client transcript replay and workflow continuation retain their separate consumers and horizons. A display heading such as “Confirmed constraints” adds no verification force and does not classify an opaque model checkpoint: MEM-OBJ-memory-display is a readable persistence view distinct from get selection.

Requested get delivery is pull. The automatic same-ID todo supply is push, with an actual filename identity match selecting delivered plan text. The overflow summarizer receives a code-selected oldest chunk and prior summary without a foreground request, so that supply is coarse push. Corpus lexical ranking occurs on a requested route; it supports ranking authority but not inferred-lexical push. These distinctions follow MEM-RTE-tiered-context, MEM-RTE-todo-context, MEM-RTE-history-compaction and EPI-RTE-corpus-retrieve.

Consolidation describes the implemented summary transformation. The epistemic overlay leaves the exact semantic relation indeterminate without an input/output pair, which does not erase the consolidation wiring or add synthesis. Nonempty-output admission is not a fidelity check. Summary eviction is forgetting; current-version evaluation filtering withdraws current reliance while retaining older records. The latter is separate from duplicate-run rejection, which supplies a validation criterion and enforcing veto rather than deduplication of stored content: EPI-RTE-summary-retain, EPI-RTE-workflow-evaluation.

Changed capability artifacts and access metadata are retained separately from the session content they might manage. Supported active consumers can use restored or replaced artifacts; memory-class registration alone leaves instance null. No value above upgrades that branch to usable context delivery. Likewise, requested user approval and model-authored source do not establish manual admission: RT-RTE-extension-admission, MEM-RTE-memory-admission. Checkpoint validation/enforcement is branch-specific: unequal nonempty hashes reject; hashless migration can reuse cached work without verifying its original executable identity, as recorded in EPI-OBJ-checkpoint.

Reconciliation supersedes the epistemic working-memory object with MEM-OBJ-working-memory, and the epistemic corpus object with MEM-OBJ-knowledge-corpus. Those canonical objects are used here. The narrower epistemic compaction/retention and ingestion/retrieval routes remain valid parts, not superseded duplicates. No other supersession changes a value.

All ten axes retain incomplete coverage. In particular, external Astra history storage/transformations, Mem0's nonoperational adapter, feedback consumers, downstream trajectory training, initial caller control and untraced process/workspace routes cannot be filled in from names or exports. The supported summary-route yes therefore coexists with unresolved learning alternatives; no system-wide no is derived. MEM-OBJ-plugin-history, MEM-OBJ-execution-records, MEM-OBJ-conversation-events and MEM-RTE-corpus-retrieval retain those limits.

Source-read log: none. This profile uses only the supplied accepted records and reconciliation; it adds no source quotations, analyst records or new evidence.