AriGraph synthesis

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

Bounded synthesis

Evidence and operating boundary

AriGraph is a memory system whose shipped consumers use retained semantic facts and episodic text for game decisions and question answering. This account covers the repository at commit e884b76d7fa5185a3a8a55e5a67393b5a43f5ef2, with applicability cutoff 2026-10-12. It includes the game and QA drivers, maintenance prompts, alternative memory baselines and evaluators, not only the graph API. SRC-1 supplies inspected implementation, declared design, attributed performance tables and small committed game/QA log samples. No fresh target execution or causal experiment was performed. External services, exact weights and the linked paper/site are outside the frozen evidence. See the boundary and evidence and Source register.

The README claims superior game performance and scalability and reports QA comparisons. The accepted records establish memory-to-decision wiring and bounded recorded sequences, not replication of aggregate results or causal effects of individual memory mechanisms: RT-CLM-memory-performance, RT-RTE-game-cycle, RT-RTE-qa-cycle.

Observation becomes retained memory

The representative workflow is one Ariadne game attempt. An operator configures the script and supplies an API key; Python owns reset, ordering, context construction and dispatch. Separate synchronous model roles select retrieval entities, judge exploration, revise a plan and choose an action. The graph makes extraction and replacement requests. These are sequential computational handoffs, not independently scheduled delegated workers. The shipped defaults allow one attempt with 150 outer iterations, five recent steps and two selected episodes. A fresh graph is constructed per attempt. See RT-RTE-game-cycle and the workflow account.

The driver cleans environment text, adds task/step material and sends the observation and previous plan to the entity selector. Cleanup can remove text, so the retained observation is not untouched environment output. Graph admission occurs before inventory is appended to the later planner's observation. The extraction model proposes short facts; code parses, normalizes and suppresses duplicates. Malformed fragments can be omitted. Requests for concrete facts and hypothesis marking are prompt policies, not source-fidelity checks: EPI-RTE-observation-cleanup, EPI-RTE-triplet-parsing, EPI-CLM-concrete-extraction.

The main store combines symbolic triplet structure, model-interpreted fact text, original extraction-input text and dense retrieval representations. These parts have different consumers. Spatial edges feed fixed navigation rules; serialized facts and episodes feed interpreting models; numerical representations influence selection. Most retained state is process-local, not a graph service or restart-persistent checkpoint. Text logging is wired, while JSON serialization is an unused API capability. See MEM-OBJ-triplets, MEM-OBJ-obs-episodic, MEM-OBJ-graph-embeddings, MEM-OBJ-logs.

Maintenance changes availability, not warrant

For game maintenance, a model compares selected old facts with new extraction outputs under conservative same-aspect replacement instructions. Code consumes the proposed old targets and separately adds new facts. Exact membership checks reject nonexistent targets; deletion protects facts whose two endpoints are caller-supplied locations. However, the parser discards the proposed replacement side. It does not establish that a supported replacement exists before deleting an old fact. The model judgment therefore has operational withdrawal force without independent acceptance of its truth: EPI-CLM-conservative-replacement, EPI-RTE-removal-parsing, RT-RTE-game-memory.

Withdrawal is not synchronized across memory. Removing a live fact and its fact vector leaves the episode text and earlier extraction associations available. A later model can receive historical text containing a withdrawn fact without an explicit stale marker. Plans and snapshots are not rewritten either. The implementation provides no semantic-error transaction or rollback; attempt reset is a different lifecycle operation. See MEM-OBJ-triplets, MEM-OBJ-obs-episodic, MEM-OBJ-episodic-snapshots.

Acquired input warrant remains unknown. Extraction, summarization, spatial inference and answering have indeterminate semantic relationships to their inputs in the assessed evidence. Syntax checks and relevance ranking do not establish entailment or truth. The inspected memory updates establish candidate retention and use, not accepted ampliative claims or post-acceptance integration: epistemic bounded conclusion.

Selected memory reaches decisions and effects

The driver requests retained facts and episodes. Model-generated entities and depths drive recursive semantic search, with top six and a 0.75 threshold per expanded query, not a global six-fact or token ceiling. Game episode ranking combines overlap with the previous selected subgraph and similarity to the old plan. Top-k selection precedes an exact-string recent-observation filter with no backfill. Stored episode keys lack the inventory appended to recent observations, so that filter does not guarantee suppression of the recent window. See RT-RTE-game-memory and its memory annotation.

Selected fact strings and episode texts enter planner/action user messages as informational context. The whole planner response replaces the previous plan and guides later retrieval, exploration, planning and action selection. Delivery establishes an available behavior-shaping path, not that each memory item changed a decision. See RT-BAP-memory-decisions, RT-OBJ-plan, RT-BAP-plan-decisions.

The action parser falls back to look on JSON/key failure, but does not enforce membership in the supplied action list or validate the complete action schema. TextWorld owns actual command semantics. A synthetic go to action can expand into several directional commands before another planning request; only the final feedback reaches the next ordinary memory update. Navigation diagnostics can also become the next observation without a distinct failure object. Thus the outer iteration budget bounds neither command count nor expenditure. Completion or budget exhaustion ends the attempt, with logs rather than an explicit returned success object. These recovery paths are wired, not demonstrated reliable recovery: RT-RTE-graph-navigation, RT-BAP-action-dispatch, workflow account.

QA and alternate routes change the assessment

QA resets memory per question, extracts facts from paragraphs without game-style deletion, retrieves with question-derived entities and ranks episodes using the freshly retrieved subgraph. Facts and paragraph texts reach the answer model. Reference answers score output afterward; they do not approve facts or trigger an installed correction. The supplied QA sample records a nonexact answer. HotpotQA loader/consumer shapes disagree, and the non-LLaMA setup selects the graph-model identifier for answering rather than the separately supplied QA identifier. These conditions prevent treating all declared QA configurations as working equivalents: RT-RTE-qa-cycle.

QA metrics license lexical comparisons under the script's normalization and counting rules, not explanation validity or source truth. Repeated tokens expose a difference between set-valued overlap counts and list lengths. The intended fact-replacement evaluator has executable defects and no established functioning scored-validation route. A separate fixed-reference diagnostic has no established consequential consumer: EPI-RTE-answer-metrics, RT-RTE-graph-evaluation, EPI-RTE-reference-match.

Full-history, summary and smart-RAG are separately invoked baseline scripts, not interchangeable backends selected by the main driver. They respectively supply complete history, model-replaced compressed memory, or ranked observation/action pairs. Summary fidelity is unchecked. Smart-RAG combines recency, retained model importance and embedding similarity; its recent-entry expression includes the oldest entry because it starts at index zero. Hypergraph offers thesis/event retention and removal with ascending event-score selection, but no inspected shipped caller establishes downstream planner use. The auxiliary ObservedGraph likewise offers mutation/display without an established later agent consumer: RT-RTE-full-history, RT-RTE-summary-cycle, RT-RTE-smart-rag, RT-RTE-hypergraph-update, MEM-RTE-observed-update.

Supported adaptation and unresolved improvement

The strongest supported contribution is an automatic experience-to-retained-state-to-later-consumer pathway. Game facts, revised plans, baseline summaries and smart-RAG metadata can shape later decision context within an attempt. This supports wired trace-fed adaptation, independently of model-weight training or cross-restart persistence. It does not establish improved capacity for future action, content-directed criticism of a consumed theory, or exercised self-improvement: MEM-OBJ-triplets, RT-OBJ-plan, RT-RTE-summary-cycle, MEM-OBJ-rag-metadata, profile verification.

The runtime records do not provide separate theory-builder condition 1–4 assessments. Retained fact/plan text is not sufficient to assert a formulated explanatory theory; delivery is not demonstrated theory-dependent consumption; no content-directed criticism and retained criticism-driven next round are established. These remain unestablished rather than system-wide evidenced absences. Reflection is likewise not established by subject matter or plan revision alone. Computational ownership of the game loop does establish operation without per-action human approval, but not an autonomous theory builder. The named reflexion prompt has no wired invocation in the searched consumers; this bounded absence does not negate the positive adaptation paths or settle opaque model reasoning: RT-RTE-game-cycle, RT-ABS-reflex-invocation, epistemic bounded conclusion.

For the task-success objective over one attempt, retained updates enter live context and selection paths. The evidence does not establish subsequent behavior causally depending on a specific improvement-directed change; exercised self-improvement and successful improvement therefore remain unestablished. No inspected driver admits changes to source code, static prompt templates or model weights. Remote provider internals and exact externally resolved checkpoints remain unresolved: RT-RTE-game-memory, RT-BAP-memory-decisions, RT-CMP-gpt-agent, RT-CMP-retriever, RT-CMP-llama-agent.

For an attempt-local game, the system supplies historical context and revisable live facts, but not a validated current-world view or guaranteed episode completion. For paragraph QA, it supplies task-local retrieval and diagnostic scoring, not metric-driven memory improvement. A pinned full-episode trace linking an update, selected content and ensuing behavior would strengthen activation findings. A controlled comparison varying a memory mechanism would change component-effect attribution. Checked replacement linkage and synchronized stale-content handling would change maintenance guarantees; repaired and exercised evaluators would change the validation account. Restart serialization plus a shipped reload consumer would change the retention horizon. None of those changes is established by the current records.

Limitations

Limitation	Affected SRC IDs and record citations	Inspected boundary	Conclusion prevented	Evidence that would resolve it
Unresolved alternate spatial-store inventory, retained from reconciliation and report verification	SRC-1; RT-RTE-hypergraph-update, MEM-OBJ-hypergraph-content, MEM-OBJ-hypergraph-index, EPI-OBJ-thesises; report limits	graphs/hypergraph.py writes reciprocal movement facts into inherited triplet storage when locations change under a non-go to action; the separate retained part and its relationship to thesis/event state are unassessed.	Complete Hypergraph storage, write-lineage, maintenance and authority comparisons. The omission neither invalidates the shipped main game/QA account nor establishes deployed alternate navigation/planner use.	Separate canonical spatial-state record, its lineage and lifecycle, and a traced consumer account relating it to thesis/event state.
Grouped profile units	SRC-1; MEM-OBJ-episodic-snapshots, MEM-OBJ-summary, RT-BAP-plan-decisions; profile limits	Snapshot text/association/score/tensor parts, summary decision/update consumers, and plan interpretation/update/ranking paths are grouped. Notes preserve supported values and local bases.	Fine-grained inventory counts and one-to-one part/consumer comparisons; not the positive forms and effects themselves.	Separate units for each operative part and consumer path with corresponding accepted records.
Bounded recorded operation and no causal comparison	SRC-1; RT-RTE-game-cycle, RT-RTE-qa-cycle, RT-CLM-memory-performance	Initial committed game/QA samples and README tables; no fresh execution, full-episode trace or intervention in this analysis.	Current-default reproduction, exercised recovery, aggregate replication, memory-item activation, improved capacity and component-level causal benefit.	Configuration-pinned complete traces and objective-relative controlled comparisons with independently varied components.
Semantic fidelity and acceptance unresolved	SRC-1; EPI-RTE-observation-cleanup, EPI-RTE-triplet-parsing, EPI-RTE-removal-parsing, MEM-OBJ-obs-episodic	Cleaned input, syntax admission, model replacement and later historical text delivery in inspected game/QA updates.	Validated fact truth, faithful extraction, warranted replacement and universal stale-content withdrawal or accepted-claim integration.	Output-to-source semantic checks, explicit scoped evidence-consuming acceptance, checked replacement transactions and later-consumer stale-content traces.
Diagnostic evaluation and QA configuration defects	SRC-1; RT-RTE-qa-cycle, EPI-RTE-answer-metrics, RT-RTE-graph-evaluation	QA loader/model selection and lexical scoring; defective intended removal evaluator. Dataset/task contents not exhaustively assessed.	Equivalent working QA branches, standard metric interpretation in all cases, functioning scored replacement validation and metric-fed production improvement.	Repaired configuration/scoring/evaluator paths with pinned exercised results; a traced consequential consumer for any claimed improvement feedback.
External models and auxiliary consumers unresolved	SRC-1; RT-CMP-gpt-agent, RT-CMP-llama-agent, RT-CMP-retriever, MEM-RTE-observed-update, MEM-ABS-state-readers, MEM-OBJ-logs	Inspected adapters and shipped callers; provider internals, exact upstream weights, generic auxiliary caller control and external log/snapshot/JSON use remain unresolved.	Provider-retention/training conclusions, exact model reproducibility, auxiliary deployed authority and graph restart reuse.	Frozen model/service identities and inspectable relevant state contracts; specified auxiliary principals, payloads and shipped later-consumer or reload traces.