GMemory synthesis

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

Bounded synthesis

From benchmark request to remembered context

GMemory implements a cross-task adaptation pathway: completed agent tasks feed retained examples and automatically revised guidance that later agents can consume. This is its strongest supported contribution. Whether those changes improve future action remains unresolved; no authenticated execution or causal comparison establishes that result (RT-CLM-evolving-memory, RT-RTE-insight-maintenance).

The evidence is source implementation and declared design from SRC-1, https://github.com/bingreeky/GMemory, frozen at commit 7b581c51d993bd600df14691d101d7e601040cc6, with an analysis cutoff of 2026-10-15. The whole-system functional boundary includes the shipped runner, AutoGen, DyLAN and MacNet consumers, memory maintenance, prompts, persistence and environment adapters. Comparator memories and external provider/simulator internals are not assessed as GMemory behavior. No target program, test, example or provider call was run. Available log headers do not authenticate GMemory mode or the generating revision; they do not establish observed operation (boundary, report verification).

An operator selects a benchmark, model, workflow and explicitly selected g-memory mode. A fixed Python module map builds one memory object and shared chat callable before iterating dataset tasks. This is a bounded benchmark curriculum, not an open-ended assistant or a remote-worker service. Python schedules own next steps; models propose textual actions, rankings, critiques and memory edits. Actions reach the selected benchmark wrapper under the launching process's authority. FEVER Search can access Wikipedia. Fresh model calls and agent names do not establish OS or provider isolation (RT-RTE-autogen-actions, RT-RTE-dylan-actions, RT-RTE-macnet-actions, runtime account).

At task start, the schedule requests remembered material. GMemory combines task-name similarity, graph neighborhoods, label filtering and model relevance ranking to select successful examples. Fallback searches can replace graph-expanded candidates; the graph need not contribute to every return. Rules are selected using positive task-name association counts, not their scalar survival scores. All three schedules discard returned failed examples, though failures remain inputs to comparative maintenance. Optional role projection rewrites selected rules before delivery and can silently produce an empty list. The ordinary CLI defaults request one successful example, zero failed examples and three insights; these are count budgets, not token limits (RT-RTE-memory-delivery, MEM-OBJ-task-layer, MEM-OBJ-insights-metadata, EPI-OBJ-projected-insights, memory read-back).

Selected task descriptions, key steps, detailed trajectories and rules enter agent user messages as advice. The full current task trajectory is concatenated into subsequent prompts; summarize does not compact it. Each agent adapter creates a fresh system/user pair rather than retaining hidden conversation history. Thus separate call frames restrict what is passed between invocations, but do not remove the task history explicitly copied into them. Model interpretation, not an enforced action filter, determines whether guidance affects a proposal (RT-BAP-retained-guidance, MEM-OBJ-mas-message, workflow account).

Acting, checking and delivering

One ALFWorld/AutoGen task is the representative implemented workflow, not a reproduced episode. The solver proposes an action. If its normalized action equals the two preceding actions, code substitutes an answer from the ground_truth persona. That helper receives role-selected memory and the existing trajectory, not an explicitly appended unexecuted current proposal. It is another model persona on the shared backend, not a correctness oracle. Repetition can occur for a valid action; a varying wrong action need not trigger substitution. The selected text reaches the environment, whose observations and rewards return to the next prompt (RT-RTE-autogen-actions, workflow account).

Outcome checks have benchmark-local meanings. ALFWorld uses simulator done; PDDL uses wrapper reward/done. FEVER compares a normalized Finish label with the dataset answer; an incorrect Finish terminates the episode but is still failure in final feedback. ScienceWorld latches configured regular-expression matches in observations. These checks supply consequential task labels, not validation of a model's explanation or generalized strategy. FEVER additionally imports Wikipedia passages: the adapter keeps URL metadata internally without enforcing its display in the returned observation. Acquired text is not knowledge produced by GMemory, and its truth is not checked by import (EPI-RTE-fever-check, EPI-RTE-subgoal-check, EPI-RTE-wiki-acquisition, RT-RTE-autogen-actions).

AutoGen ends environment interaction on done or outer-step exhaustion, then saves success or failure and calls backward before returning the reward/success tuple. The runner records the result only after that synchronous memory work completes. Memory failure can therefore block runner-visible delivery after environment completion. Separate provider-attempt limits and outer-step limits do not guarantee task termination: empty answers retry without advancing the workflow counter. There is no automatic task restart in this selected schedule (workflow account, RT-CLM-bounded-recovery).

The other consumers change coordination and feedback, not just labels. DyLAN traverses a layered candidate network with optional critiques, consensus, model-ranked elimination and a final decision model; it resets local caches per step and saves without post-task backward. MacNet traverses a spatial graph and forwards temporal predecessor outputs. Its reused node histories are not cleared by task-context reset, so latest outputs can cross tasks when temporal edges exist. This is a conditional implemented path, not observed contamination or benefit, and it is separate from restart-persistent GMemory (RT-RTE-dylan-actions, RT-RTE-macnet-actions, MEM-OBJ-node-memory).

Turning completed work into later guidance

Task retention stores readable task content and derived extras in Chroma, with interaction graphs serialized in document metadata. A separate NetworkX task-query graph persists as a pickle; current rules and their metadata persist in JSON. The query graph is not the recorded interaction graph, and insight associations are dictionary lists rather than an interchangeable third graph. These stores reload on startup, but their writes are not one transaction (MEM-OBJ-mas-message, MEM-OBJ-state-chain, MEM-OBJ-task-layer, MEM-OBJ-insights-rule, RT-RTE-task-retention).

Save-time processing attempts to delete negative-reward states, but pops a temporary list slice and leaves backing states unchanged. Successful stored trajectory reconstruction therefore precedes, and is distinct from, digit erasure in clean_traj. That cleaned text feeds model key-step extraction and failure diagnosis; erasure can collapse truth-relevant numeric distinctions. Loaded interaction-graph reconstruction also hides its last completed state because the public view excludes an assumed trailing open state that deserialization does not add. This does not prove loss of separately stored display trajectory text. Neither model extraction nor diagnosis has a semantic quality gate (EPI-RTE-trajectory-cleanup, MEM-RTE-state-roundtrip, EPI-OBJ-key-steps, EPI-OBJ-fail-reason).

By default, rule maintenance begins at five saved tasks and repeats every fifth task. Software samples successes/failures and supplies existing rules; models propose ADD, EDIT, REMOVE or AGREE operations. Shape, punctuation, banned-word, index and duplicate checks control admission, not rule truth. AGREE resolves text again after numeric checking; a lookup miss returns -1 and can credit the last rule. REMOVE lowers survival score rather than necessarily withdrawing the rule immediately. Positive associations rank selection, while positive scalar scores permit survival. AutoGen/MacNet outcome credit adjusts substring-matched cached raw rules by +1 or -2 without checking whether the delivered/projected rule was followed (RT-RTE-insight-maintenance, EPI-RTE-rule-disposition, MEM-OBJ-insights-metadata).

Every twentieth saved task separately triggers clustered rule merging. This model numbered-list path bypasses ordinary operation checks, clears/rebuilds rules and resets associations and scores. Its requested reduction count is not an enforced cap. Extraction and merging afford consolidation; actual semantic preservation remains unresolved. Rule replacement and score decay are implemented, but requested generality does not establish newly produced, warranted claims. Causal clauses can survive in rules, yet ordinary solving does not replay the separate maintenance criticism logs or per-output justifications (MEM-RTE-cluster-merge, MEM-OBJ-insights-rule, profile verification).

What the adaptation establishes

The task-to-derived-extras-to-rules-to-later-consumer chain supports a wired trace-fed memory update, including durable behavior-shaping text. It does not require model-weight training, and the name finetune_insights does not establish such training. At solving consumers the text supplies advice; at update consumers retained experience and rules feed durable revision. Those forces are distinct from ranking and survival controls. Maintenance automatically supplies selected retained inputs to new update invocations, and MacNet automatically forwards temporal output; these push operations coexist with requested task-start retrieval, rather than making that pull chain unsolicited delivery (RT-RTE-task-retention, RT-RTE-insight-maintenance, MEM-OBJ-node-memory, profile verification).

For later benchmark success, the architecture supplies revisable guidance and outcome-responsive survival feedback. Improved capacity for future action, particularly improvement caused by criticism of a consumed theory, is not established. Model key steps, diagnoses, generalized rules, merged rules and projected variants have indeterminate semantic relations without concrete outputs. Syntax admission, persistence, selection and use are not evidence-consuming acceptance of a general claim against a named truth criterion and scope; no post-acceptance integration of such claims is established in these inspected routines. An exposed output-improvement judgment prompt has no found consequential caller in the recorded search (RT-RTE-insight-maintenance, epistemic bounded conclusion, EPI-CLM-output-improvement).

The independent property assessments remain separate:

  • Theory-builder condition 1, localized content: afforded. Individual rules can state and replace whole principles; authenticated retained theory contents were not supplied.
  • Condition 2, consumption: afforded. Models receive rules for interpretation, but actual decisions depending on their content are unobserved.
  • Condition 3, content-directed criticism: claimed. Prompts request criticism, but no required challenged claim, result, placement of blame and revision rationale establishes qualifying criticism.
  • Condition 4, iteration: afforded. Revised rules persist for later rounds; a qualifying criticism result shaping the next round is unobserved.
  • Reflection: representations of performance and strategies feed updates and can return to solving; actual two-direction behavioral dependence remains unobserved. A reflective theory builder is not established.
  • Autonomy: computation performs internal sampling, proposal calls, parsing and writes without per-rule human approval. This automation is supported; it does not establish an autonomous theory builder while the independent conditions remain unresolved.
  • Self-improvement: over the shipped cross-task horizon, source wiring supports improvement-directed self-change toward the independently specifiable objective of later task success. Retained guidance can enter the system's own later prompt organization. Authenticated subsequent operation causally depending on the change is missing, preventing an exercised self-improvement finding; successful improvement separately needs outcome evidence.

These assessments rest on RT-RTE-insight-maintenance and the epistemic bounded conclusion, not a single system grade. Rule-level addressability and reload persistence do not establish their finer semantic or causal properties.

The assessment would change with authenticated traces linking task evidence, admitted rule versions, selected/projected text and subsequent decisions, plus outcome comparisons at the granularity claimed. A memory-package intervention could support package-level attribution; isolating a rule or component requires varying it independently. Concrete challenged claims, blame placement, revisions and later reliance could resolve the criticism/iteration questions. Semantic acceptance criteria and checked transformations would change the warrant account; versioned transactional persistence, explicit inner-loop deadlines and corrected serialization would change the corresponding operational guarantees, but those changes are not present evidence (RT-CLM-evolving-memory, RT-RTE-insight-maintenance, MEM-RTE-state-roundtrip, RT-CLM-bounded-recovery).

Limitations

Limitation Affected SRC IDs and record citations Inspected boundary Conclusion prevented Evidence that would resolve it
Stale attribution retained as the reconciliation's unresolved conflict and report verifier's limit SRC-1; RT-RTE-task-retention, MEM-RTE-state-roundtrip, memory integration issue 1, report-verification limit Current runtime retention record and memory annotation/correction request Memory still attributes effective reward pruning to runtime, but current runtime explicitly agrees deletion is ineffective. Neither a continuing mechanism disagreement nor a necessary current runtime correction follows. An owner-corrected memory attribution consistent with the pinned runtime; no further mechanism evidence is needed to describe the current agreement.
No authenticated operation or causal benefit evidence SRC-1; RT-CLM-evolving-memory, RT-RTE-insight-maintenance, report verification Source control flow, prompts and selected log headers Exercised adaptation, actual dependence on memory and improved capacity or component benefit Revision/configuration-authenticated route traces and scoped interventions with outcomes and confounding limits.
Generated semantics and criticism remain unresolved SRC-1; EPI-OBJ-key-steps, EPI-OBJ-fail-reason, EPI-OBJ-projected-insights, RT-RTE-insight-maintenance Extraction, diagnosis, rule generation, projection and later delivery Faithful compression, checked derivation, semantic novelty, accepted generalized truth and fulfilled theory-builder conditions Linked inputs/outputs, semantic checks, explicit criticism and revision evidence, and later content-dependent reliance.
Lossy retained views have unmeasured practical consequences SRC-1; EPI-RTE-trajectory-cleanup, MEM-RTE-state-roundtrip Slice/pop, digit erasure and graph serialization implementation Universal meaning preservation, full graph roundtrip fidelity or measured downstream error Concrete retained views and consumer traces; corrected transformations with roundtrip and semantic checks.
Initialization and embedding provenance unknown SRC-1; MEM-OBJ-mas-message, MEM-OBJ-task-layer, MEM-OBJ-insights-rule, RT-CMP-embedding-func Loaders and known subsequent writes Complete lineage or admission control for existing Chroma, pickle, JSON and vectors Store snapshots with author/admission history and pinned vector-generation provenance.
External models, dependencies and resources are not independently captured SRC-1; RT-CMP-gpt-chat, RT-CMP-embedding-func, boundary Client wrappers, configuration and environment adapters Immutable model identity, provider-side learning/retention, retrieval quality, simulator correctness or present reproducibility Pinned external artifacts/contracts and inspectable scoped dependency/provider evidence.
Outcome and acquisition licenses are narrow SRC-1; EPI-RTE-fever-check, EPI-RTE-subgoal-check, EPI-RTE-wiki-acquisition Reference equality, observation patterns and paragraph import General strategy truth, causal diagnosis, source truth or complete displayed lineage Audited references/pattern fidelity, source-version lineage and claim-specific evidence evaluation.
No end-to-end recovery or persistence guarantee SRC-1; RT-CLM-bounded-recovery, RT-RTE-task-retention, MEM-RTE-cluster-merge, workflow account Inner retries, synchronous save/backward, separate store writes and destructive merge Finite task duration, coherent atomic restart, automatic rollback or guaranteed delivery after environment completion Explicit deadlines, coordinated/versioned writes and inspected failure/recovery traces.
Consumer-dependent alternate activation SRC-1; RT-RTE-dylan-actions, RT-RTE-macnet-actions, EPI-OBJ-projected-insights DyLAN save, MacNet histories, role/trajectory projection Uniform backward feedback, actual temporal-edge effect or shipped use of trajectory-conditioned projection Actual graph/configuration and consumer traces; a concrete caller for the optional projection branch.
Data, static prompts and simulator internals are not exhaustively assessed SRC-1; boundary coverage, report verification Declared functional coverage, selected supporting source reads Dataset coverage/split integrity, contamination, exhaustive prompt quality or simulator correctness Dedicated inspection of the relevant datasets, prompts and simulator semantics under a declared evidence boundary.

No source reads were made for this synthesis; its findings use the supplied members and their retained evidence passages.