Agno memory profile

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

Comparison rationale

Boundary, coverage and evidence strength

This revision-2 profile classifies accepted records, not a deployed Agno instance. Every axis is partial because the included inventory retains opaque/custom alternatives or missing initial-control, payload or later-consumer facts. A known built-in unit means complete classification only within its explicitly named operation or part. It does not certify all adapters or branches. The inventory limits attached to MEM-OBJ-agent-session, MEM-RTE-learning-context, MEM-RTE-knowledge-context and WF-RTE-sequential-delivery prevent whole-runtime memory completeness. No observation or causal evidence supports stronger bases than wiring; documented external/manual consumer roles remain afforded.

Component versions changed and subsequently consumed through RT-RTE-config-admission are retained behavior-shaping state. Static cookbook prompts, skills and registered executable code are not accumulated memory. Work-queue storage is not substituted for this memory boundary. EPI-RTE-sft-selection contributes retained exported examples and score-consuming admission, but not an established trainer or updated model weights. Provider inference components describe numerical implementation, not evidence that use updates those weights.

Forms, provenance and forces remain local

Dense knowledge embeddings are parametric representations in vector storage; symbolic IDs/metadata and natural-language document text remain separate parts. MEM-OBJ-learned-knowledge does not make embedding provenance fully known. MEM-OBJ-result-store likewise distinguishes opaque payload, access index and preview. Classifying its readable preview does not classify the entire preserved string. Filesystem knowledge and opaque media/provider state retain uncertainty rather than receiving values from adjacent inspected branches.

Preferences and organizational directives can supply instruction, while facts, evidence and continuity supply knowledge. Their system-message placement alone does not determine the force. Existing conversations/context additionally supply learning at extraction/summary update consumers, independently of downstream knowledge or instruction. MEM-BAP-retained-context, EPI-RTE-learning-extraction and EPI-RTE-learning-recall support these different paths. Retained text/vectors/aliases affect retrieval ranking even without a learned ranking policy. Entity retirement markers enforce ordinary-display withdrawal; their historical retention makes that curation invalidation rather than erasure. ResultStore scope fields validate a read and enforce refusal where the recorded check applies. These controls do not imply universal authorization.

Admission and transformation

Software/model admission is automatic even when an operator starts extraction, maintenance, sync or export. Direct API/HTTP edits and initial caller content remain not-determinable for write agency: generic caller identity does not establish a human decision or its absence. MEM-RTE-memory-editing and RT-RTE-config-admission retain this distinction. PROPOSE asks for approval but the writer checks no approval token; the request cannot establish an enforced manual admission finding for MEM-RTE-learned-knowledge.

Summaries and standing-fact extraction consolidate retained conversation content even when originals remain. Model fidelity is unobserved; a dropped qualification does not establish synthesis. LearnedKnowledge guidance requests generalization but no accepted record establishes a concrete new-claim transformation, so synthesize remains unresolved rather than asserted. MEM-RTE-profile-curation establishes conditional age/count decay and normalized-string deduplication, not semantic equivalence. Index rebuilding, schema checks and score-controlled SFT export are not by themselves promotion or synthesis. Page revision and pruning, entity historical withdrawal and direct row revision keep their distinct curation warrants.

Selection chains and qualifying trace updates

Automatic host supply is push; requested searches, archive lookup and result reads are pull. A result identifier on a requested payload does not create identifier-based push. On automatic routes, session/user/entity/namespace/config matches select delivered retained parts; enablement, recency and budgets are coarse selectors. Current-message semantic search supports inferred-embedding; entity term relevance supports inferred-lexical. A supersession judgment changes writes, not push selection. No inferred-judgment push is inferred from opaque reranking or instructions to apply relevant content. See MEM-RTE-learning-context, MEM-RTE-entity-memory, MEM-RTE-knowledge-context and RT-RTE-context-window.

Automatic facts/profile/continuity/insight extraction and retained session/compaction summaries have trace-fed write-to-later-consumer chains. They support trace_learning yes without proving improved performance, novel knowledge or a complete final-outcome snapshot. Summary-only and transcript-bearing folds both qualify where retained/replayed; unsupported archive storage does not create durable recoverability. MEM-RTE-tool-compression does not establish complete durable provider replay. Raw offload/session logs and exported SFT examples alone do not establish a closed derived behavior-shaping/training update. No bounded absence records establish trace_learning no, and no pull-only or nonqualifying local route makes the partial trace-source inventory inapplicable.

Session-log input is supported on conversational extraction. Compaction history additionally carries tool-call/result traces; that value is local to the evidenced tool-bearing history route, not a claim that concurrent learning snapshots contain final tools. Imported references, static instructions and opaque content are not automatically reclassified as activity traces. Neither stored outcome text nor ordered turns alone establishes trajectories. MEM-RTE-session-summary, MEM-OBJ-compaction-archive and RT-RTE-context-window retain the relevant input distinctions.

Canonicalization and source-read log

The supplied reconciliation supersedes the epistemic duplicate learning object with MEM-OBJ-learned-knowledge, and the epistemic Curator route with MEM-RTE-profile-curation. All profile references use those canonical declarations. The conditional EPI-OBJ-profile-memories remains a valid part, not another whole profile. Function-level epistemic extraction/save/recall records remain valid beside the broad memory route. No supersession changes an evidence basis, provenance uncertainty, curation meaning or consumer finding.

Source reads: none. Classification used only the supplied accepted members and reconciliation; no ambiguity required a frozen-source reread and no evidence or records were added.