GMemory boundary

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

Boundary and evidence

GMemory is in scope as a memory/knowledge/context-engineering system. The target is the G-Memory implementation together with its shipped task runner, three multi-agent consumers, prompt construction, persistence and maintenance paths, and environment feedback and recording adapters. Its intended use is to carry task experience and distilled guidance into later multi-agent tasks. This is a whole-system functional boundary, not an analysis limited to the memory API. The in-repository task runner and workflows are included callers, not an assumed external host. Alternative memory implementations are comparator modes, not additional target systems (SRC-1, README.md, tasks/run.py, mas/module_map.py).

The code-grounded classification rests on the implemented task-to-memory-to-task loop. The runner constructs one memory module before iterating over tasks. The workflows retrieve retained trajectories and insights, format them into agent prompts, record interactions and environment outcomes, and submit completed contexts to memory. GMemory stores task records in Chroma, persists and reloads a task graph and insight JSON, and invokes model calls for trajectory extraction, relevance ranking, insight revision and role-specific guidance (SRC-1, tasks/run.py, tasks/mas_workflow/autogen/autogen.py, tasks/mas_workflow/dylan/dylan.py, tasks/mas_workflow/macnet/graph_mas.py, tasks/mas_workflow/format.py, mas/memory/mas_memory/GMemory.py, mas/memory/mas_memory/memory_base.py). These are source paths, not a claim that execution was reproduced.

The shipped prompt consumption is explicit:

user_prompt: str = format_task_prompt_with_insights( few_shots=few_shots, memory_few_shots=successful_shots, insights=roles_rules.get(solver.profile, raw_rules), task_description=self.meta_memory.summarize() ) --- tasks/mas_workflow/autogen/autogen.py

The same consumer closes the feedback loop:

self.meta_memory.save_task_context(label=final_done, feedback=final_feedback) self.meta_memory.backward(final_done) --- tasks/mas_workflow/autogen/autogen.py

Reload is part of the implementation rather than an inferred service responsibility:

if os.path.exists(self._graph_save_path): with open(self._graph_save_path, 'rb') as f: self.graph = pickle.load(f) --- mas/memory/mas_memory/GMemory.py

self.persist_file: str = os.path.join(self.working_dir,f'{self.namespace}.json') self.insights_memory: list[dict] = load_json(self.persist_file) or [] --- mas/memory/mas_memory/GMemory.py

Maintenance includes model-mediated rule operations, score-based pruning and periodic merging. The prompts are part of that mechanism, not instructions governing this analysis (SRC-1, mas/memory/mas_memory/GMemory.py, mas/memory/mas_memory/prompt.py):

AGREE : REMOVE : EDIT : ADD: --- mas/memory/mas_memory/prompt.py

Consumer differences remain inside the boundary. In particular, AutoGen and MacNet call backward after saving the task; the inspected DyLAN schedule saves the task without that call. A later account must not assume identical maintenance across all three consumers (SRC-1, tasks/mas_workflow/autogen/autogen.py, tasks/mas_workflow/macnet/graph_mas.py, tasks/mas_workflow/dylan/dylan.py).

Frozen evidence and dependencies

The sole frozen source is SRC-1 at full commit 7b581c51d993bd600df14691d101d7e601040cc6; the acquired checkout's HEAD was verified against that pin. The analysis cutoff is 2026-10-15, the supplied run date. Only this revision supplies evidence: the linked paper, upstream benchmark repositories, model weights, live pages and provider services have not been independently captured. The local checkout is a read-only access root, not the source identity.

The overall evidence tier is code-grounded. Source text establishes implemented control flow and declared intent, not successful execution or measured benefit. No target program, test or example was run. External dependencies include an OpenAI-compatible chat endpoint and credentials, SentenceTransformer embeddings, Chroma, NetworkX, FINCH, and benchmark packages/resources for ALFWorld, ScienceWorld and PDDL; FEVER uses a Wikipedia tool adapter. The shipped wrappers, configuration and feedback interpretation are inside the boundary; external service behavior and third-party internals are not attributed to GMemory (SRC-1, mas/llm.py, mas/utils.py, requirements.txt, tasks/envs/alfworld_env.py, tasks/envs/sciworld_env.py, tasks/envs/pddl_env/pddl_env.py, tasks/envs/fever_env.py). This prevents conclusions about provider reliability, embedding quality, simulator correctness or present-day reproducibility from this source alone.

Top-level coverage

All rows refer to SRC-1. Included paths remain eligible for further inspection at the same commit; initial inspection is not an evidence allowlist. “Not yet inspected” below denotes available source, not an access gap.

Source register