ReasoningBank synthesis
Type: agentic-system-analyses/types/agentic-system-synthesis.md
Bounded synthesis
ReasoningBank implements an experience-to-advice-to-task loop. Its strongest supported contribution is automatic production of durable guidance from successful and failed task trajectories, followed by selection and delivery to later task agents. This is a wired trace-learning architecture, not evidence of improved future capacity. No execution trace or controlled comparison establishes actual reliance, transfer or benefit. The repository's learning and memory-aware scaling statements are design claims, not experimental results within this analysis (RT-RTE-web-admission, RT-RTE-swe-admission, RT-BAP-system-memory, EPI-CLM-memory-learning, EPI-CLM-memory-scaling).
The evidence is SRC-1 implementation and shipped instructions from https://github.com/google-research/reasoning-bank at ed80611788292ea739f1effd31f16c53823b8a0d, with cutoff 2026-10-14. The whole-system boundary includes the WebArena and modified mini-swe-agent callers, producers, evaluators, persistence and later consumers, not merely a retrieval API. Provider internals, live benchmark environments and installed external host behavior are not captured. Whole-system scope does not imply exhaustive inspection of every adapter or a working invocation for every option. This synthesis uses the accepted members only; it adds no source reads. See the boundary and Source register.
A new task receives selected prior advice
An operator supplies benchmark tasks and run bounds. The WebArena driver schedules inference, evaluation and induction subprocesses in order; the coding batch driver schedules dataset instances through worker threads. Code owns scheduling and retrieval. Models choose browser actions or shell commands, while BrowserGym or the selected shell environment owns execution effects (RT-RTE-web-pipeline, RT-RTE-swe-task).
Before task execution, ordinary memory-enabled callers reload a local JSONL bank and a separate query-embedding cache. They append the current query to the cache but rank the previously loaded snapshot. Instruction-aware similarity over prior query vectors selects one task ID; an identity join takes the first matching bank row and delivers all its advice blocks. Selection ranks queries, not advice truth, task success or evidential quality. An unmatched top ID produces no advice rather than a search through lower-ranked IDs. A one-record budget is neither a one-item nor a token budget (MEM-OBJ-query-cache, MEM-RTE-web-retrieval, MEM-RTE-coding-retrieval).
The web runner overwrites a delivery text file that the agent rereads each step. Coding puts the selected string into a fresh task's system message. At that consumer, retained prose supplies model-mediated directions with relevance discretion, not a symbolic permission rule. The runner requests memory from its selector; automatic task setup supplies the result to a task model that did not itself request memory. These are distinct request and supply operations. Dense vectors determine ranking; identifiers determine lookup; advice reaches the model as interpreted text. Neither placement in a system message nor successful retrieval establishes compliance (MEM-OBJ-delivery-text, RT-BAP-system-memory).
Task execution and delivery have narrower checks
The representative SWE-Bench workflow connects scheduling, iterative execution, prediction delivery and new advice within one shipped driver. A model proposes a bash block, a parser requires exactly one block, and the configured environment executes its contents. Format errors and recognized timeouts become feedback for another turn in the same context. That recovery does not verify patch correctness or undo prior shell effects. Docker is the default executor, but local host-shell and imported implementations are selectable. Prompt restrictions on edits are not parser-enforced safety rules (RT-RTE-swe-task, WF-RTE-action-recovery, RT-RTE-client-selection).
A submission marker terminates the task protocol; it is not benchmark acceptance. The worker saves the trajectory and writes the returned result as a prediction before judging or inducing advice, even when that result is an error rather than a submitted patch. Retrieval occurs before the execution recovery block. A later judging or append failure can leave a prediction without new advice, and ordinary restart skips existing prediction keys instead of resuming the missing admission stage. The shipped launcher uses one worker; the multiworker option does not extend the prediction lock to bank/cache writes (RT-RTE-swe-task, RT-RTE-swe-admission, workflow delivery and continuation annotations).
Working-context controls are also local rather than universal guarantees. Web code shrinks a reconstructed observation/history view before adding system memory; Claude-prefixed names bypass fitting and iteration exhaustion can leave an oversized prompt. Coding observations retain head and tail of long shell output; fuller material requires an explicit later read. Separate task, judge and generator contexts exchange selected text and returned labels/advice, not a restorable context pointer. These controls do not produce durable semantic summaries. Coding agent query budgets exclude direct judging, induction and embedding calls (RT-RTE-context-control, RT-RTE-swe-admission).
Feedback chooses how to generate advice, not whether its claims are warranted
The ordinary web evaluator judges task execution against completeness, grounding and the right target using a bounded saved view, including at most five page captions. It has no expected-answer input in that route. Its verdict and available explanation select and inform success/failure induction. The gt alternate instead uses saved benchmark reward. Vendored exact/inclusion, URL and HTML checks establish configured predicates conditional on extraction and reference fidelity; model-backed fuzzy checks have separate fallibility. None licenses an induced explanation or its future applicability (EPI-RTE-web-check, EPI-RTE-benchmark-check).
The coding judge receives the task and saved non-system transcript, not an external SWE-Bench test result. Its substring parser treats any text containing success as success, including unsuccessful as a static counterexample. Either label selects an induction prompt; failure does not veto new advice. Coding also omits the previously selected system advice from the generator's input and passes no judge rationale through its Boolean interface. This prevents interpreting the chain as explicit criticism and revision of that prior advice (EPI-RTE-swe-check, RT-RTE-swe-admission, MEM-OBJ-feedback).
Induction asks for reusable procedures and failure-prevention lessons. Host code splits generated text on blank lines and appends it without a semantic item gate. Requested item counts, generalization and non-overlap are not validated. Web rows retain extracted thought/action arrays beside the prose; coding rows retain query, ID and outcome while the transcript remains separate. Later task delivery selects advice only, not those supporting artifacts. Reasons might survive in generated prose, but their preservation and later use are not guaranteed (MEM-OBJ-web-advice, MEM-OBJ-coding-advice, MEM-OBJ-lineage-ledger).
Reducing retained trajectories into advice supports trace consolidation, not demonstrated revision of old advice or creation of warranted new propositions. No generated item is supplied, so its actual content could be reshaping, derivation, conjecture or policy-only guidance. Task evaluation precedes that content and does not constitute evidence-consuming acceptance of it for a named future use. The assessed append/selection chains retain and use candidates without establishing post-acceptance integration. This is bounded to those chains, not a negative about every unassessed command (EPI-OBJ-web-memory, EPI-OBJ-swe-memory, epistemic authority ledger, profile verification).
What persists, and what remains to be demonstrated
The ordinary loops retain advice and access structures across problems without updating task-model or embedding-model parameters. Exact provider weights and historical vector versions are not pinned. Query-cache admission is independent of advice admission, so failed induction can leave unmatched keys. Redo or task-output deletion does not withdraw prior bank/cache rows; duplicate IDs still select the first bank match. These are source-level consequences, not observed incidents (RT-CMP-task-model, RT-CMP-gemini-embedding, RT-CMP-qwen-embedding, MEM-OBJ-query-cache, MEM-OBJ-trajectory).
For the ordinary induction routes, theory-builder condition 1, localized formulation, is afforded by the item instructions. Conditions 2, content-dependent decisions, 3, claim-directed criticism with revision or changed reliance, and 4, criticism shaping a next round, are each uninspected. Criticism of a consumed theory improving future capacity is separately unestablished. Reflection has a wired trajectory-to-representation-to-agent pathway, but its actual representation-mediated behavioral effect is uninspected. Scheduling and admission are computational within operator bounds; autonomous theory-building remains uninspected because every required internal operation is not established. These are independent conclusions, not a single grade (RT-RTE-web-admission, RT-RTE-swe-admission).
At the task-to-task horizon, retained guidance affords dispositional improvement-directed self-change: task experience can shape material subsequently supplied through a live advisory channel. Exercised self-improvement is not established because subsequent behavior causally depending on that change is unobserved. Successful improvement additionally needs outcome evidence. Fixed model weights do not negate this architectural route, and automatic persistence does not prove its exercise (RT-RTE-web-admission, RT-RTE-swe-admission, RT-BAP-system-memory).
The scaling branch does not establish a working distinct-trajectory comparison: induction repeatedly reads the final trial directory, lacks model forwarding, calls an uninstantiated client class and leaves the ordinary tuple return incompatible with string-block delivery. Its inverted status variable is not an explicit label delivered to the generator. Standalone induction has unresolved import/launch context and no established later destination consumer. Standalone regrading has a missing import binding and no automatic bank-revision consumer. Thus a bounded sequential advice-supply scenario has stronger implementation support than the claimed memory/scaling synergy; neither has observed benefit here (RT-RTE-web-scaling, scaling annotation, MEM-RTE-standalone-induction, EPI-RTE-answer-regrading).
A trace linking generated advice to later decisions would change the activation assessment. Claim-specific criticism retained into a revised next round would change the relevant theory-builder assessments. Controlled memory/no-memory comparisons would support conclusions at that treatment level; attributing a component effect requires independent variation of that component. Scaling needs repaired distinct-input, client and payload wiring plus execution evidence before its intended comparison can be credited. These requirements do not erase the already supported ordinary trace-fed supply mechanism (epistemic claim comparison, RT-RTE-web-scaling).
Limitations
Neither accepted verification declares an additional limit, and reconciliation reports no unresolved conflict. The following evidence boundaries still prevent stronger conclusions; none would incorrectly imply no remaining limitation.
| Limitation | Affected SRC IDs and record citations | Inspected boundary | Conclusion prevented | Evidence that would resolve it |
|---|---|---|---|---|
| No generated advice, execution trace or causal comparison | SRC-1; RT-RTE-web-admission, RT-RTE-swe-admission, EPI-OBJ-web-memory, EPI-OBJ-swe-memory | Ordinary shipped producers and later supply routes; implementation/design only | Actual advice semantics, compliance, criticism-shaped iteration, improved capacity, exercised self-improvement and causal benefit | Linked generated items, evaluation and downstream decision traces; comparisons with the necessary independent variation |
| External model and environment state is uncaptured | SRC-1; RT-CMP-task-model, RT-CMP-gemini-embedding, RT-CMP-qwen-embedding, EPI-RTE-benchmark-check | Shipped clients, executors and vendored predicates, not live services or installed BrowserGym behavior | Exact reproducibility, deployed isolation, reference fidelity and agreement between model feedback and benchmark outcomes | Frozen provider/environment identities, host integration traces and validated benchmark references |
| Initial banks, external edits and historical vector provenance are unresolved | SRC-1; MEM-OBJ-web-advice, MEM-OBJ-coding-advice, MEM-OBJ-query-cache | Known subsequent automatic writes and selector consumption | Complete content lineage, initial admission control, authenticity and vector compatibility | Initialization/edit records and per-vector producer/version lineage |
| Feedback and admission do not validate generated advice | SRC-1; EPI-RTE-web-check, EPI-RTE-swe-check, RT-BAP-system-memory | Ordinary trajectory evaluation, append and advice-only delivery | Truth, causal explanation or transfer warranted merely by a task verdict, retention or selection | Advice-specific evidence-consuming tests and scoped acceptance records; retained reasons linked to later reliance |
| Alternative production and delivery chains remain incomplete | SRC-1; RT-RTE-web-scaling, MEM-RTE-standalone-induction, EPI-RTE-answer-regrading | Scaling wiring/return contract, standalone imports/output routing and regrading import | Working distinct-trajectory synthesis, scaling synergy, standalone later learning or automatic regrading-based revision | Repaired invocation and payload contracts, distinct trial provenance and linked downstream execution |
| Recovery, concurrency and context guarantees are local | SRC-1; RT-RTE-swe-task, RT-RTE-swe-admission, RT-RTE-context-control, WF-RTE-action-recovery | Traced task drivers, parser/timeout recovery, bank/cache writes and context controls | Transactional memory-loop completion, universal budget/safety guarantees, rollback or complete restoration | Coordinated recovery/write protocols, advice-inclusive budget enforcement and deployment-specific executor evidence |
| Ancillary routes are not exhaustively inspected | SRC-1; RT-RTE-client-selection, RT-RTE-context-control, MEM-OBJ-trajectory; report verification, profile verification | Selected shipped drivers and materially traced alternatives, not every agent/adapter | Complete route inventory, all-entry-point equivalence or system-wide absence of acceptance, withdrawal or other memory mechanisms | Targeted inspection of remaining material consumers and admissions with records preserving their distinct scope |