post-graph-rag memory profile

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

Comparison rationale

Boundary and coverage

The profile classifies the accepted records, not fresh source observations. PostgreSQL graph payloads and vector columns are retained memory; process-local query parameters and static shipped prompts are not. File retention is included for shipped evaluator results and replay inputs. Custom proposal provenance, possible extra storage and unassessed evaluator consumers remain local unresolved units rather than being erased by inspected database witnesses. See MEM-OBJ-documents, MEM-OBJ-embeddings, MEM-RTE-evaluation-memory and RT-RTE-extractor-extension.

Known units describe complete classification only within their named part and consumer path. All positive bases are implementation wiring except the separately afforded inspection consumer. No finding claims observed execution, reliable model compliance or causal benefit. The profile makes no complete-value assertion for unresolved injected or evaluation-adoption alternatives. Their unknowns are not evidence that extra behavior exists.

Encoding, lineage and authority

Graph storage and symbolic form are independent: stored prose has natural-language consumption, metadata has fixed-rule consumption, and learned dense vectors have distributed-parametric consumption. The deterministic term-hash fallback is symbolic rather than learned. Provider training provenance remains unresolved without obscuring the known call that produces a retained embedding. See MEM-OBJ-embeddings and RT-CMP-local-embedding.

Imported source passages and software/model-derived products coexist. Session-derived facts and retrieval-derived archive fields additionally have trace-extracted lineage. No generic caller identity proves human-authored memory or per-content human admission. The known automatic write mechanisms remain classified even where initial proposal provenance is unknown. See RT-RTE-document-admission, MEM-OBJ-documents and MEM-RTE-evaluation-memory.

Retained prose supplies knowledge at model/context consumers. Embeddings, lexical text and ratings influence ranking. Stored eligibility fields are validation criteria with enforced exclusions on their named consumers; routing fields select operation targets or input scope. These are distinct from the static prompts' instruction force. No all-channel withholding is inferred from archive, dormancy or supersession: old passages and reports can still carry an assertion. See MEM-BAP-access-state, MEM-OBJ-relation-metadata, MEM-OBJ-community-metadata and EPI-RTE-supersession-write.

Transformations and later consumption

Extraction and reporting consolidate material. Model requests and lossy projections do not establish faithful semantic preservation, but that uncertainty does not turn consolidation into synthesis. No recorded witness establishes a newly introduced claim as the implemented synthesis operation. Same-key/alias merges both deduplicate and revise existing payloads. Supersession/withdrawal invalidate reliance while retaining graph residue, and retention demotes access eligibility. See MEM-OBJ-entities, MEM-OBJ-community-text, EPI-RTE-relation-upsert, EPI-RTE-community-render and MEM-RTE-document-withdrawal.

Questions request corpus retrieval; builds request graph/child-report input; maintenance requests event/statistical input; replay requests prior baseline records. Selection and delivery inside those chains remain pull, not another push merely because software chooses results. Therefore the assessed pull chains have no push signal. Unknown additional consumers prevent a whole-inventory inapplicable signal assessment. See RT-RTE-question-answer, RT-RTE-community-build, RT-RTE-retention-control, MEM-RTE-change-polling and MEM-RTE-evaluation-memory.

Trace-fed writes and unions

Two wired routes qualify independently. Retrieval events feed archive writes whose flags alter later seeds/snapshots. Session turns feed automatic durable fact construction whose results reach later benchmark questions. Their update inputs have learning authority; archive flags downstream enforce eligibility and extracted facts downstream provide knowledge. The original trace sources are respectively event-streams and session-logs. Neither route requires parameter training, a new true proposition, observed improvement or own-agent experience. Instance-local cleanup bounds conversational persistence but does not remove later-invocation consumption. See MEM-OBJ-retrieval-events, RT-RTE-retention-control and MEM-RTE-evaluation-memory.

The derived learning union supports yes at wired basis, with partial inventory coverage. No no is inferred from parameter fixity, saved grades without an update consumer, or raw logs alone. Other evaluated/injected origins and adoption mechanisms remain unresolved. Session wrappers do not add an event-stream source, and neither hash records nor session chronology alone establishes tool-traces or trajectories. See EPI-RTE-benchmark-check and RT-RTE-extractor-extension.

Supersessions and source-read log

Reconciliation replaces the duplicate epistemic passage with MEM-OBJ-documents, report prose with MEM-OBJ-community-text, and source-withdrawal route with MEM-RTE-document-withdrawal. This profile uses those canonical records. The replacements alter no controlled value: they identify the same retained objects/routes while preserving their limits. Entity/relation containers remain valid alongside their separately consumed parts, and community rating remains a part of the broader metadata inventory rather than a duplicate.

No source files were read. All classifications and assessment warrants come from the supplied accepted members and their record passages; no records, quotations or evidence were added.