Tortoise memory profile

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

Comparison rationale

This comparison classifies the accepted records, not the repository afresh. All positive findings are implementation-level wiring except the explicitly afforded diary read. There is no observed-operation or causal-benefit upgrade. Each axis remains partial because the included writer, maintenance and consumer inventory is incomplete, not because implementation evidence is intrinsically too weak to classify a route.

Objects, derivation and control

Graph content, local backlog and optional journal/record projections are distinct retained surfaces. Record-capable session writes and graph-only source expansion do not establish restoration or retrieval parity. The relevant warrants are MEM-OBJ-session-turns, RT-OBJ-event-emission and MEM-RTE-source-expansion. Transient payload dictionaries and process-local numerical work are not additional accumulated-memory substrates.

Dense embeddings and fixed-rule numerical belief fields have different forms. The former encode distributed numerical representations; the latter have localized fields consumed by fixed code. Neither is a change to extraction-model weights. Known encoding/maintenance paths do not settle all initial provenance: RT-CMP-embedding-model, MEM-OBJ-belief-state and EPI-OBJ-belief-state.

Authored checkpoint/diary lineage does not establish manual admission. A generic caller may be human or computational. Conversely, starting capture does not make the ensuing software/model admission manual. These distinctions follow MEM-RTE-checkpoint-diary, RT-RTE-direct-point and RT-RTE-session-capture.

Authority, transformation and trace learning

Conversation experience feeds a durable graph update, and retained graph priors enter gap review before that update. Those are learning-authority paths at the update consumer; selected output has knowledge authority at the later host model. They do not imply improved capacity or reliable factual acceptance: MEM-OBJ-derived-points, WF-RTE-gap-review and RT-BAP-framework-block.

The qualifying v2 route supports trace-learning yes and conversational session-log origin. Raw spool retention alone does not establish another qualifying learning route. Neither graph Event envelopes nor a model call establish additional original trace types: MEM-OBJ-capture-spool, MEM-OBJ-session-turns and MEM-OBJ-derived-points. No accepted evidenced-absence record warrants a local trace-learning no; opaque alternatives remain unresolved rather than negative.

Checkpoint duplicate rejection, re-capture revision, tombstone withdrawal and correction promotion have identifiable implemented effects. Extracting standing graph content from retained session messages also supports wired consolidation: bounded/redacted turns pass through narrative, mapping and review into durable claims. This operation does not require proof of semantic faithfulness. The epistemic account's indeterminate truth-apt relation leaves novel claims and factual warrant unresolved, not consolidation unsupported. Possible extraction errors do not establish synthesis. The extraction unit remains partial for unresolved additional transformations, and opaque alternative extractors remain unresolved. Numerical maintenance is assessed separately: dream recomputation is not automatically decay or semantic revision. See MEM-OBJ-derived-points, RT-RTE-session-capture, WF-RTE-gap-review, EPI-OBJ-extracted-points, and the other transformation records: MEM-RTE-checkpoint-diary, MEM-OBJ-session-turns, MEM-RTE-point-retraction, EPI-RTE-review-integration and MEM-RTE-dream-writeback.

Delivery and coverage

Automatic callback supply is push at the host boundary, while the memory service fulfills a request inside that chain. Source-session expansion is requested evaluation retrieval, not independent push or established live reflex delivery. The session-start digest's recency selection supports coarse push; the canonical pipeline's unresolved search internals prevent choosing lexical, embedding or judgment matching merely from its API name. Printed pointer IDs do not establish identifier matching: RT-RTE-context-delivery, MEM-RTE-hook-volunteer, MEM-RTE-start-digest and MEM-RTE-source-expansion.

Lifecycle enforcement is consumer-specific. Canonical terminal exclusion does not cover the recent-session digest, whose inspected query lacks that predicate. No union of these findings means universal withdrawal. Likewise unknown complete remote commit consumption, spool producers and caller control prevent extrapolating one wired path across all admissions: MEM-RTE-point-retraction, MEM-RTE-start-digest, RT-RTE-production-commit and MEM-OBJ-capture-spool.

The supplied reconciliation supersedes the duplicate epistemic hook route with MEM-RTE-hook-volunteer, and the duplicate epistemic withdrawal route with MEM-RTE-point-retraction. This profile uses those canonical destinations. Their supplemental advisory-channel and message-invalidation findings do not create extra routes or stronger universal guarantees. Payload versus admitted content, overlapping belief fields and maintenance versus retention functions remain distinct; no other supersession changes a value.

Source-read log: none. No source ambiguity required inspection beyond the supplied accepted members.