AI Agents in Depth memory profile

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

Comparison rationale

This is a partial comparison of separate workflows, not a shared suite runtime. All positive bases are implementation wiring, not observed operation or demonstrated benefit. The unresolved alternatives in every axis preserve the accepted memory inventory; no directory name supplies a storage or learning value. There are no supersessions: reconciliation preserves the context container and its epistemic parts and finds distinct identities rather than replacements.

JSON is not a behavioral form. Facts/backstories are natural-language information at model consumers, while addresses, resume controls and executable policy have symbolic consumers. The PDF and currency parts of RT-OBJ-conversation-history retain their different derivation paths through EPI-RTE-pdf-import and EPI-RTE-currency-conversion. Shipped model weights and static instructions do not become accumulated memory merely because they participate in inference.

Update and downstream forces differ. MEM-RTE-trace-maintenance and MEM-RTE-sequential-campaign consume dialogue to produce durable context; the derived output supplies knowledge to later answering calls. The trace-learning positives identify this automatic write chain, not improved capacity, accepted truth or parametric training. Initial files, direct-helper caller control and provider payload provenance remain unresolved. No bounded absence records justify a trace-learning no.

The deterministic consolidation in MEM-RTE-offline-maintenance reduces existing claims, deduplicates normalized equal text and forgets superseded/excess notes. It does not validate contradiction, synthesize new claims or retain an invalidation history. Interactive replacements support evolve. Arbitrary sequential JSON output does not establish every requested semantic transformation. Retry-policy generation in RT-RTE-policy-candidate creates retained candidate source from the read-only shipped stable input identified by RT-OBJ-retry-policy. That static input is excluded from accumulated memory; changing its rule while creating a candidate does not establish evolve over an existing retained memory entry. Retention of candidates and reuse of rejected feedback do not establish a controlled curation operation over those objects. This branch therefore has no positive curation finding and remains not-determinable, without withdrawing interactive evolve or retry-policy trace learning.

The pull CLI and automatic context supply coexist. MEM-RTE-conversation-reload selects all facts coarsely but identity-filters active-session history; recent processing additionally matches processed turn IDs. MEM-RTE-sequential-campaign uses checkpoint identity/signature/purpose to reuse retained work and selects latest memory wholesale. A requested query return is not a second push. MEM-RTE-history-search remains a capability with no established shipped model consumer; semantic-search/reranking settings do not establish vector storage, parametric payload form or end-to-end read-back.

RT-RTE-policy-candidate supplies a separate trace-fed behavior-shaping artifact: retained candidate source is consumed by evaluator function calls, while rejected reasons reach a later coding request. Synthetic trajectories and rejected evaluator feedback retain their respective trajectory and tool-trace origins. This structural yes does not install a successor or close a self-improvement attribution chain. Candidate enforcement is bounded to RT-BAP-candidate-policy; eligibility and artifact delivery do not imply stable deployment.

Source-read log: no source files were read. Classification used only the supplied accepted members, their governing types/contracts and reconciliation. Missing recorded facts were retained as coverage limits, not filled by new evidence.