EXIA SEED boundary

Type: agentic-system-analyses/types/agentic-system-boundary.md

Boundary and evidence

This run examines EXIA SEED as a memory/knowledge/context-engineering system: its shipped conversational-memory plugin, maintenance tools, prompts and evaluation drivers. The intended use is to explain how conversations and explicit notes become retained, cited subject pages and how later queries consume that state. It is not an analysis of the enclosing Hermès assistant runtime. The boundary is whole-system for EXIA SEED's shipped artifact, including its callers and evaluators, rather than just its retrieval API (SRC-1, README.md, plugin/exia_seed/__init__.py, plugin/exia_seed/outils/consolider.py, banc/locomo/preparer.py, banc/locomo/repondre.py).

The target qualifies because its consolidation issues model calls to produce retained content, and its provider serves that content back to a model-driven host. The host-facing registration and consolidation adapter establish this boundary in implementation (SRC-1):

def register(ctx) -> None: ctx.register_memory_provider(ExiaSeedProvider()) --- plugin/exia_seed/__init__.py

rep = call_llm(task="exia_seed_consolidation", provider=modele_cfg.get("provider") or None, model=modele or modele_cfg.get("default") or None, messages=messages, reasoning_config=parse_reasoning_effort(effort), timeout=timeout) --- plugin/exia_seed/consolidation.py

The included functional loop covers provider initialization and configuration; primary-conversation ingestion and explicit-note/native-memory admission; privacy and secret handling; batch closure; model prompting and deterministic fragment checks; page, index, provenance and graph maintenance; automatic recall and explicit search/read tools; and the shipped replay and scoring paths. The provider's shipped subprocess caller and nightly shell driver are included, not treated as external merely because they call the consolidation API. The loop was traced through sync_turn, _lancer_consolidation, _consolider, prefetch, _souvenirs and handle_tool_call in SRC-1 plugin/exia_seed/__init__.py; the consolidation CLI in plugin/exia_seed/outils/consolider.py; page admission and progress recording in plugin/exia_seed/consolidation.py; and state reload in plugin/exia_seed/rappel.py. The latter reconstructs its memory cache when the graph timestamp changes. These are source-level paths, not observations of execution.

Admission and downstream consumption are both material to the boundary. The checker rejects agent-only cited material, while tool descriptions tell the host model how to search retained content (SRC-1):

if que_agent: return False, "repose seulement sur des paroles de l'agent" --- plugin/exia_seed/consolidation.py

"Search your long-term memory (EXIA SEED) when the INDEX.md summary is not enough. Give a SHORT " --- plugin/exia_seed/langues/en.py

The frozen evidence is the acquired Git checkout at c7497ac96da20bd1818096e35a59b6c0237f9730, verified against its HEAD. The applicability cutoff is 2026-11-07. The overall tier is code-grounded; documentation supplies declared intent and attributed operation, not independently observed deployment. Version labels do not replace the commit pin: the manifest says 0.9.1 while the changelog describes further changes (SRC-1, plugin/exia_seed/plugin.yaml, CHANGELOG.md).

Shipped-path coverage

The tracked tree has the following top-level entries. All included paths remain available for later inspection at this commit. “Not yet inspected” means available source detail, not an access failure. Table roles are grounded in SRC-1 at the named paths; the initial inspection is deliberately shallower than the later mechanism analysis.

External dependencies and prevented conclusions

Hermès provides the MemoryProvider contract, hook dispatch, agent context, threading helper, configuration/home/logging facilities and auxiliary model client. Its source is not frozen in this run. Host-side scheduling, context placement, tool selection, authentication renewal and final assistant behavior therefore cannot be established independently; only EXIA SEED's side of those interfaces can be analyzed. The Docker image named by the test and benchmark launchers is also not captured, so loader compatibility and reproducible execution in that image remain unverified (SRC-1, plugin/exia_seed/__init__.py, plugin/exia_seed/consolidation.py, tests/banc.sh, banc/locomo/lancer.sh). Provider services and model weights are external; source inspection cannot establish model quality, latency, cost or successful calls.

Private Nexa memories, private benchmark panels, generated run caches and full public benchmark datasets are not supplied as frozen run evidence. Included drivers still permit analysis of their construction and consumption paths, but not independent reconstruction of reported scores or actual production retention. banc/modeles/consolider.py also expects a separate production-plugin copy under its benchmark root; that copy is not the acquired source and cannot be silently substituted for it. The internal specification named in the plugin is expressly unpublished, preventing verification against that complete private contract (SRC-1, README.md, banc/modeles/consolider.py, plugin/exia_seed/config.py).

Public examples are not literal production data. The source's warning says (SRC-1):

*English summary: names of people, companies and places, and private information, are fictitious in this public version, but based on real cases. --- docs/00-AVERTISSEMENT.md

This prevents treating substituted names, prices or sensitive examples as original deployment traces. No observed-run or causal-experiment layer is established by this boundary; later use of target-authored measurements must preserve their attributed status and stated conditions. References to upstream systems, protocols and papers do not register those repositories or live pages as evidence.

Source register