xMemory boundary

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

Boundary and evidence

This run analyses xMemory as a memory/knowledge/context-engineering system, not as a general agent runtime. Its intended use is retaining conversation information and retrieving it for later questions. The whole-system boundary includes the Python library and the shipped LoCoMo drivers: those drivers coordinate hierarchy maintenance, reload retained content and turn retrieved memories into model answers. Restricting the target to its public facade would omit material responsibilities shipped in this repository (SRC-1, xMemory/__init__.py, src/api/facade.py, evaluation/locomo/add.py, evaluation/locomo/xMemory_search_framework.py).

The classification is code-grounded. The facade delegates message ingestion, episode flushing, semantic completion and search to the memory system. Episode creation invokes a generator and saves its output; prompts define model-dependent extraction. These are memory operations serving subsequent model calls, not merely a static document collection (SRC-1, src/api/facade.py, src/core/memory_system.py, src/generation/prompts.py).

def add_messages(self, user_id: str, messages: Sequence[Dict[str, Any]]) -> Dict[str, Any]: """Append messages to the memory buffer for a given user.""" return self._memory_system.add_messages(user_id, list(messages)) --- src/api/facade.py

You are an episodic memory generation expert. Please convert the following conversation into an episodic memory. --- src/generation/prompts.py

The included loop covers input formatting and admission, buffering and episode boundaries, episode and semantic generation, prediction/correction extraction, persistence and indexing, theme split/merge maintenance, graph export, baseline and adaptive retrieval, original-message hydration, answer generation and evaluation. This is a boundary declaration, not a finding that all these paths execute successfully. In particular, the construction driver explicitly coordinates themes and graph updates; these responsibilities are not assigned to an excluded host (SRC-1, evaluation/locomo/add.py, src/core/memory_system.py, src/core/memory_hierarchy.py).

new_eps, new_sems = memory.fetch_incremental_batches(user_id)

    result["stage"] = "updating_themes"
    theme_info = memory.update_themes(user_id, new_sems)

--- evaluation/locomo/add.py

The retrieval driver is likewise included, rather than treated as an external example. It reloads persisted hierarchy files and supplies retrieved content to its answer model (SRC-1, evaluation/locomo/xMemory_search_framework.py).

storage_root = Path(self.config.storage_path) themes_path = storage_root / "themes" / f"{user_id}_themes.jsonl" --- evaluation/locomo/xMemory_search_framework.py

memories = self.search(memory, user_id, question) self._hydrate_episode_results(memory, user_id, memories.get("episodic", [])) self._hydrate_semantic_results(memories.get("semantic", [])) response = self.answer(question, memories) --- evaluation/locomo/xMemory_search_framework.py

Tracked top-level coverage

The tracked tree has six top-level entries. All functional shipped paths are included. “Not yet inspected” below describes initial inspection depth, not an access gap or a functional exclusion. Later jobs may inspect any file at the same pin under SRC-1.

All coverage statements refer to SRC-1 at the frozen commit. The inspected construction, persistence, maintenance, reload and answer-consumption paths establish the whole-system scope; the table does not claim a whole-repository line-by-line review.

Dependencies, exclusions and evidence limits

External dependencies include model weights and inference libraries, embedding-provider services, Chroma and local filesystem storage. The source configuration names a Llama model, an embedding model and filesystem/vector backends; the utility export selects the local-model adapter, while the embedding adapter contains OpenAI/Azure client paths (SRC-1, src/config.py, src/utils/__init__.py, src/utils/embedding_client.py, environment.yml). Their external implementations, live service behavior and downloaded model artifacts are outside this frozen source. This prevents claims about provider reliability, model quality or exact reproducibility from repository text alone.

The paper, project website, linked upstream systems, external PerLTQA dataset and release-hosted prebuilt memories are not frozen sources. They are excluded even when linked by the README; this prevents verification of their contents, paper-wide results, upstream behavior and prebuilt-memory quality. The bundled LoCoMo input and shipped evaluation code remain included (SRC-1, README.md, evaluation/locomo/evals.py). No live host application outside this repository is included; the run cannot characterize deployment integration or autonomous agent behavior beyond the shipped memory/QA workflow.

The applicability cutoff is 2026-10-15. The evidence is the acquired Git revision 1a4f9dd8abd7f2da86bfe6afe4293a1ea1c63608, not a moving branch or the README's publication dates. No execution traces, completed result artifacts or causal experiments were supplied as frozen evidence. Source inspection supports implementation and declared design; it does not establish successful execution, benchmark gains or operational durability. Available files left for later inspection are not access failures.

Source register