EXIA SEED synthesis
Type: agentic-system-analyses/types/agentic-system-synthesis.md
Bounded synthesis
From conversation to retained memory
EXIA SEED is a conversational-memory plugin for the Hermès assistant. Its distinguishing architecture puts a model at the write side and fixed lexical/graph computation at the production retrieval side: it turns selected conversation lines and explicit notes into cited subject pages, then supplies selected page content to later assistant invocations. This code-grounded account covers the shipped plugin, maintenance tools and evaluation callers at commit c7497ac96da20bd1818096e35a59b6c0237f9730, with an applicability cutoff of 2026-11-07. SRC-1 supplies implementation, design statements and reported operation, but no registered execution trace or causal experiment. Hermès internals, provider internals and actual deployment state are outside the frozen evidence. Thus this is an account of implemented capacities, not independently established performance or successful assistant behavior (boundary, source register).
The representative connected workflow is capture, batch closure, asynchronous consolidation and later recall. Hermès owns the conversational turn and final assistant action. EXIA SEED's primary-context hook masks recognizable secrets, handles retention commands and appends speaker-labelled source lines. Day, volume and session triggers close batches. A host-context thread launches the shipped consolidation CLI; an already busy thread does not queue another job. The consolidator takes a nonblocking advisory lock and discovers pending files. The auxiliary model receives numbered source lines, node/alias and concept catalogues, and instructions—not the foreground assistant's full context or an action executor. Code owns scheduling, admission and writes (RT-RTE-conversation-ingest, RT-RTE-page-consolidation, workflow account).
Primary-only capture does not mean primary-only knowledge admission. Explicit-note tools and native-memory mirrors can append register entries in every initialized context, and consolidation marks those entries as non-agent sources regardless of the caller. This is a separate trust route, not proof of human confirmation. Guarded callers reject recognizable secrets, but direct register and benchmark raw-file writes do not inherit guards they bypass. Neither the first-batch budget exception nor notes/catalogues added outside that check permit a hard total-model-input ceiling (RT-RTE-register-admission, MEM-OBJ-registre, RT-RTE-page-consolidation).
What admission establishes
The model proposes facts, destinations, citations and metadata. Deterministic code rejects defined source, speaker, number, proper-name and word-overlap failures. This veto has real operational force, but it does not test entailment or source truth. Invalid references in a mixed list can survive when valid cited text suffices; mixed user/assistant citations do not prove confirmation of the particular claim. A static counterexample in the records shows that a negative preference can support admission of its contradictory positive form after stopword removal. That establishes a lexical blind spot, not an observed generated error or its frequency (RT-CLM-citation-control, EPI-RTE-fragment-check). Calendar checking can strip unsupported dates without rejecting a fact; its derivational warrant concerns its explicit calendar predicate, not whether an event occurred (EPI-RTE-time-check).
Admitted prose is appended to Markdown subject pages, separately from fixed-parsed provenance, temporal and retrieval controls. Normalized duplicate suppression is not contradiction resolution. Existing page bodies are not supplied wholesale to fact extraction, so consolidation is not comprehensive semantic truth maintenance. Summaries are a second model-derived representation: the producer receives up to 3,000 fact characters per page in section order and stops at the first oversized fact, whereas its lexical checker can inspect the full page. Passing name/number/path/length checks therefore establishes neither whole-page model exposure nor semantic equivalence. Keywords have their own node/alias membership gate (MEM-OBJ-page-facts, MEM-OBJ-page-metadata, MEM-OBJ-summary, MEM-OBJ-keywords, EPI-RTE-summary-check). These are operational admissions for later use, not established production and acceptance of new ampliative knowledge.
Completion is partial-progress rather than transactional. Unreadable JSON leaves inputs pending; a parsed response can advance all selected batches and notes despite rejected fragments. Rejection reasons remain audit material rather than a wired repair prompt. Summary failure can preserve old values; graph or Git failure does not undo page writes. Timeout can interrupt after progress, and a zero CLI exit need not mean every fragment was admitted (RT-RTE-page-consolidation, MEM-OBJ-journals, workflow annotations).
From retained state to later context
Pages, summaries, thesaurus and graph persist in files; best-effort Git supplies history. The retrieval cache uses process-local structures and SQLite FTS. Graph weights are localized symbolic association state, not learned dense parameters. Derived concept vocabulary and keywords influence later ranking; an edge's supporting snippet licenses an association under the construction rule, not the truth of a proposition (MEM-CMP-memoire, MEM-OBJ-thesaurus, MEM-OBJ-graphe).
A later primary query triggers automatic lexical/alias/concept/graph selection. Phase 1 computes and journals but returns empty; phase 2 supplies selected subjects, summaries and facts without a foreground memory-tool request. Explicit search/read tools remain available in either phase and other initialized contexts. The cache reload signal is graph-edge modification time, so page-only or summary-only maintenance can leave an existing cache stale. Phase-2 summaries and pointers are added after generic selection; its estimated token ceiling is not a universal final-output cap. Auxiliary model calls keep unrelated foreground history out, but the compression hook itself only records a marker: host compaction and checkpoint restoration are not EXIA SEED mechanisms established here (RT-RTE-memory-delivery, RT-RTE-state-maintenance, memory read-back).
Controls remain consumer-specific. Automatic recall excludes future utterances and expired prospective facts. Search excludes future facts but not expired intentions; page-read neighbor recovery and fallback do not apply the future-date predicate. These branches are reachable through ordinary reads when neighboring material exists, but no affected deployment instance is supplied. Privacy labels are not authorization checks in these readers. Withdrawal excludes raw lines from new consolidation and selected graph references, without retracting already retained page prose. A later model receives selected prose, not all retained source comments or reasons. Returned conduct rules have interpretive instruction force, not executable enforcement (RT-RTE-memory-delivery, EPI-RTE-withdrawal-read, RT-BAP-recalled-content).
Learning, evaluation and the assessment boundary
The strongest supported adaptive contribution is the automatic session-to-page-to-later-consumer chain, including downstream summaries and retrieval-vocabulary updates. It creates durable behavior-shaping material without changing model weights. User conduct corrections can become retained procedural guidance and receive delivery priority. This supports a wired trace-fed update capacity, not demonstrated improved capacity for future action. Actual host placement, compliance and causal dependence on the update remain unobserved (MEM-OBJ-page-facts, MEM-OBJ-summary, MEM-OBJ-thesaurus, RT-BAP-recalled-content).
The supplied accounts do not establish a formulated-theory criticism/iteration route in the inspected production bodies. They do not establish that criticism of a consumed theory improved future capacity. They also do not independently establish reflection's two causal directions, autonomous theory building across every internal operation, or exercised self-improvement: later operation causally depending on retained changes is missing. These are separate unresolved properties, not consequences of one another or a denial of the trace-fed capacity. The runtime does not supply a condition-by-condition theory-builder assessment from which to infer conditions 1–4; page labels and inaccessible model processing cannot fill that gap (RT-BAP-recalled-content, epistemic bounded conclusion).
For a scenario requiring inspectable, cited conversational context on later requests, the implementation connects the necessary producer, retained store and consumer. For a scenario requiring verified factual entailment, reliable erasure, confidential-output enforcement or observed behavioral improvement, those same routes do not supply the required guarantees. Benchmark judging is separate: replay constructs memories and answers, and reference-answer judges produce reports, not live acceptance or installation decisions. LongMemEval's type-specific rubric and substring parser further bound what a correctness label means. No supplied runs establish scores or component effects. Production retrieval's non-embedding characterization must not be extended to all evaluators: report verification identifies a separate embedding baseline and page-fusion transformation missing from the detailed analyst inventory (RT-RTE-benchmark-replay, EPI-RTE-lme-judging, report verification).
Provenance-linked inputs, proposals, gate results, admitted state and later host contexts would distinguish implemented delivery from exercised use. A controlled comparison with subsequent behavior would be needed to establish benefit or attribute it to a particular update mechanism. Semantic admission checks, downstream retraction propagation, privacy enforcement at consumers and cache invalidation covering every writer would change the respective guarantee assessments; their presence cannot be inferred from citation metadata or benchmark labels alone.
Limitations
| Limitation | Affected SRC IDs and record citations | Inspected boundary | Conclusion prevented | Evidence that would resolve it |
|---|---|---|---|---|
| No supplied production execution or causal comparison | SRC-1; RT-RTE-page-consolidation, RT-BAP-recalled-content | Plugin admission and delivery implementation; external Hermès/model behavior not frozen | Actual activation, reliable compliance, improved future capacity and exercised self-improvement; no independent theory-builder, reflection or autonomy verdict follows | Provenance-linked production traces including later consumers; property-specific theory/criticism and causal evidence; controlled outcome comparisons for benefit |
| Source truth and generated semantic relations remain unresolved | SRC-1; EPI-RTE-fragment-check, EPI-RTE-summary-proposal, EPI-RTE-summary-check | Fragment lexical gates and bounded summary producer/full-page lexical checker | Entailment, faithful compression, accepted ampliative knowledge or error prevalence; the static counterexample establishes possibility only | Linked source/proposal/output instances and evidence-consuming semantic assessment for a named use |
| Unequal temporal, privacy and withdrawal controls | SRC-1; RT-RTE-memory-delivery, EPI-RTE-withdrawal-read, MEM-OBJ-page-metadata | Automatic recall, explicit search/read and graph reconstruction | Universal currentness, confidential-output confinement or erasure of derived knowledge; no actual incident established | Complete affected-state and consumer traces; downstream retraction and access-control contracts plus enforcement evidence |
| Partial progress and graph-timestamp invalidation | SRC-1; RT-RTE-page-consolidation, RT-RTE-state-maintenance, MEM-CMP-memoire | Cooperating writers, journals, summary maintenance and cache reload | Atomic recovery, complete input representation, immediate visibility after every writer or universal token caps | Fault/recovery traces, coordinated writer/invalidation protocol and final serialized-context measurements |
| Initial content, caller agency and external dependencies not frozen | SRC-1; MEM-OBJ-registre, MEM-OBJ-thesaurus, RT-CMP-consolidation-model, boundary | Known subsequent transformations, mutable model configuration and host interfaces | Complete starting provenance or human-approval history, exact provider identity, host compaction guarantees, loader compatibility or private-contract compliance | Starting-state provenance, caller/admission records, pinned host/image/provider identities and the missing contracts |
| Benchmark inventory omission carried from report verification | SRC-1; RT-RTE-benchmark-replay, MEM-OBJ-graphe, MEM-OBJ-thesaurus, EPI-OBJ-concepts, verification limit | Verifier inspected fusion and embedding-baseline paths; analyst records do not separately inventory them; other diagnostic/language detail remains incomplete | Exhaustive benchmark object/lineage/selector comparisons or extension of non-embedding production retrieval to every evaluator; unfrozen runs/data also prevent reproduced scores | Accepted detailed records for fusion/baseline parts and consumers, scoped language/diagnostic inspection, frozen datasets/run artifacts and component-isolating comparisons |
| Grouped profile units carried from profile verification | SRC-1; EPI-OBJ-evaluations, RT-RTE-benchmark-replay, MEM-OBJ-page-facts, MEM-OBJ-summary, MEM-OBJ-thesaurus, verification limit | Benchmark cached prose/judgment fields share a form unit; raw extraction and summary/vocabulary updates share a learning-authority unit | Every grouped part having every listed form or consumer effect, fine-grained part-to-consumer comparisons and route counts from grouped units; supported existence findings remain valid | Separate units tracing each operative part and actual consumer; no upgrade from partial coverage without complete inventory support |