ReasoningBank boundary

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

Boundary and evidence

ReasoningBank is in scope as a model-dependent memory system. This run covers the repository's shipped implementation, not just its retrieval API or the conceptual memory-item format. Its intended use is to derive reusable reasoning from agent trajectories and supply it to subsequent web-navigation and software-engineering tasks. The source describes that intent directly (SRC-1, README.md):

We introduce ReasoningBank, a memory mechanism for agents that learns from both successful and failed trajectories, with reasoning stored as memory content. --- README.md

The target class is memory/knowledge/context-engineering system, with a whole-system boundary. This includes the shipped callers, agents, induction prompts, retrieval and embedding-cache maintenance, trajectory persistence, feedback evaluators, and scaling variant. The WebArena driver and modified mini-swe-agent batch runner are inside the target: they own persistence and subsequent consumption, rather than being unspecified external callers. Whole-system denotes functional coverage of the shipped memory loops; it does not assert that all entry points work or that every dependency is frozen.

The initial entry-point trace establishes both sides of the memory loop:

• WebArena: WebArena/pipeline_memory.py invokes WebArena/run.py, WebArena/autoeval/evaluate_trajectory.py, and WebArena/induce_memory.py in task order. The runner reloads the JSONL bank, selects records through WebArena/memory_management.py, and writes selected content to the text file passed into the shipped agent. WebArena/agents/legacy/agent.py reads that file into the system message. Induction uses success/failure instructions and appends new records to the bank (SRC-1, these paths).
• SWE-Bench: SWE-Bench/run.sh enters third_party/ and calls mini-extra swebench. The script registration in third_party/pyproject.toml and dispatch in third_party/src/minisweagent/run/mini_extra.py lead to third_party/src/minisweagent/run/extra/swebench.py. Its process_instance reloads the bank, retrieves prior records, calls the agent with selected memory, saves a trajectory, judges that trajectory with a model, induces new items, and appends them. third_party/src/minisweagent/agents/default.py places the selected content in its system message (SRC-1, these paths).
• The WebArena scaling path is also included: WebArena/pipeline_scaling.py launches trials and invokes WebArena/induce_scaling.py, whose instructions come from WebArena/prompts/memory_instruction.py. Inclusion is not a claim that its sampling, feedback interpretation or client wiring is correct; those require mechanism-level review (SRC-1, these paths).

Minimum implementation passages supporting inclusion of the shipped producers and consumers follow. In the WebArena induction path:

# write memory to jsonl file
with open(args.output_path, 'a') as f:
    f.write(json.dumps({
        "task_id": args.task.split(".")[-1],
        "query": ex["query"],
        "think_list": ex["think_list"],
        "action_list": ex["action_list"],
        "status": ex["status"],
        "memory_items": generated_memory_item.split("\n\n"),
        "template_id": ex["template_id"]
    }) + '\n')

--- WebArena/induce_memory.py

In the modified SWE-Bench runner, retained records are reloaded before retrieval:

with open(f"./memory/{model.config.model_name}.jsonl", "r") as f:
    memory_bank = [json.loads(line) for line in f.readlines()]

--- third_party/src/minisweagent/run/extra/swebench.py

The same runner subsequently persists induced content:

    with open(f"./memory/{model.config.model_name}.jsonl", "a") as f:
        f.write(json.dumps({
            "task_id": instance_id,
            "query": task,
            "memory_items": generated_memory_item,
            "status": "success" if status else "fail",
        }) + "\n")

--- third_party/src/minisweagent/run/extra/swebench.py

The frozen evidence is the acquired Git repository at ed80611788292ea739f1effd31f16c53823b8a0d, with applicability cutoff 2026-10-14. HEAD matched that pin on inspection; the checkout reported no working-tree changes. The overall evidence tier is code-grounded. Source reading establishes implementation structure and declared instructions, not successful execution, measured improvement, or causal effectiveness. No target execution, provider call, dependency installation or experiment was performed.

Repository coverage

Every top-level tracked file or directory is classified below. “Included” makes the material available for analysis within this boundary; initial inspection is distinguished from exhaustive auditing. Paths in this table are SRC-1 commit-relative anchors.

External dependencies and conclusion limits

External provider services, credentials and model weights, BrowserGym's installed experiment runner, hosted benchmark websites/container images, SWE-Bench datasets/task repositories and the external SWE-Bench grading service are outside the frozen source. Their declared use is visible in SRC-1 README.md, pyproject.toml, WebArena/run.py and SWE-Bench/run.sh; their live behavior and exact deployment state are not captured. Consequently this analysis cannot establish live provider behavior, environment reproducibility, actual task success, agreement between model feedback and benchmark grading, or measured memory/scaling gains. Shipped adapters and vendored benchmark/evaluator modifications remain inside the boundary; responsibility for those changes is not attributed to an external host.

No external paper, linked upstream repository, live documentation page or new revision is registered. README claims may be treated as attributed design statements, not as independent evidence from its linked paper. No supplied execution trace or causal experiment is registered. Available but initially uninspected files at the frozen commit can be inspected by later analysts under SRC-1; this register is not a path allowlist.

Source register