Trajectory-informed memory generation synthesis
Type: agentic-system-analyses/types/agentic-system-synthesis.md
Bounded synthesis
Trajectory-informed memory generation describes a cross-task guidance pipeline: analyze completed agent experience, consolidate reusable tips, and select tips for the next task's initial prompt. Its strongest supported contribution is a documented experience-to-guidance mechanism accompanied by configuration-specific reported AppWorld gains. The evidence does not establish which particular lesson caused a later action or a repeated online improvement cycle. This account covers the complete arXiv 2603.10600v1 HTML capture, arxiv-2603.10600v1-html-captured-2026-10-14, with an applicability cutoff of 2026-10-14. It is document-grounded: SRC-1 supplies design and reported operation, not inspected implementation, raw execution traces or a registered causal experiment. The target is the memory system, not its enclosing agent or application APIs; the boundary is a complete artifact and a partial loop. See boundary and evidence and the Source register.
From completed experience to candidate guidance
The input is a complete trajectory containing contexts, reasoning, actions, results and optional evaluations. Model calls classify reasoning and outcomes, trace backward to decisions, and propose strategy, recovery and optimization tips. Whole-task extraction treats the trajectory as one unit. Subtask extraction first segments it into logical phases with descriptions, applications, purposes and source step ranges; independent calls then produce 2–4 tips per subtask. This is allocation of model work, not established multi-agent delegation or security isolation. The reported agent and extraction pipeline use GPT-4.1, but no exact snapshot is pinned. SRC-1 §§3.1.1–3.1.4, §4.1.2; RT-RTE-trajectory-extraction, MEM-OBJ-trajectories, RT-CMP-gpt4 and the workflow account.
Independent outcome labels are optional. Without them, extraction infers success, failure or recovery from the agent's own reflective signals. Even external reports do not make every interpretation ground truth: the paper's example converts an HTTP 400 response into a specific missing-payment diagnosis that the status alone does not entail. Likewise, backward attribution and generalized prerequisite advice can propose causal and transfer conjectures rather than merely compress a transcript. Their accuracy is not established by calling the stage causal analysis. SRC-1 §§3.1.1–3.1.3; EPI-RTE-self-outcome, EPI-RTE-report-interpretation and EPI-RTE-attribution-generation.
Consolidation changes both advice and access
Management generalizes descriptions, clusters similar operations using embeddings, and asks a model to deduplicate, resolve conflicts and combine complementary tips. Canonical descriptions replace individual per-trajectory descriptions. The model makes semantic admission decisions: successful-source guidance takes precedence over failed-source guidance, and recovery patterns over speculative prevention. This is substantive outcome-consuming curation, not storage alone. It is also best-effort precedence within detected clusters, dependent on metadata quality and model compliance; it does not independently test a merged prescription's explanation or general applicability. SRC-1 §§3.2.1–3.2.4; RT-RTE-tip-consolidation, EPI-RTE-conflict-check and EPI-RTE-conflict-disposition.
The retained product separates natural-language advice from access structures. Tips contain content, purpose, steps and triggers. Descriptions and dense vectors support matching; symbolic metadata supports filtering and precedence. Vector storage is declared, but text/metadata storage and actual persistence are unspecified. Source IDs and explanatory purpose provide intended provenance without establishing that full causal reasoning survives consolidation or can be read later. Requested synthesis also does not establish that an actual successor contains a new claim absent from its inputs. SRC-1 §§3.2–3.3; MEM-OBJ-tips, MEM-OBJ-descriptions, MEM-OBJ-embeddings, MEM-OBJ-metadata and EPI-CLM-provenance-validation. Neither new vectors nor new advice establishes model-weight updating.
Selection, delivery and execution remain separate
A new task triggers automatic supply of retained tips before agent reasoning. Cosine retrieval matches task embeddings to stored descriptions using a threshold and optional top-k. The alternative model-guided selector infers domain/category filters and constructs a hybrid query using similarity and priority, preserving generic null-context candidates. Selection grants context eligibility, not truth warrant. The model-guided evaluation requests five tips; cosine experiments include capped and uncapped alternatives. The top-3 experiment also leaves a description-count versus delivered-tip-count ambiguity, so it cannot support exact prompt-size comparisons. No token budget is established. SRC-1 §§3.3.1–3.3.2, §§4.1.4, 4.2.2–4.2.3; RT-RTE-cosine-retrieval, RT-RTE-llm-selection and report verification's limit.
Selected content, purpose, steps, triggers, category and priority enter a guidelines section after task context and before standard instructions. The later agent does not ordinarily receive the full source trajectory or attribution chain. This is advisory instruction and explanatory context, not an API permission grant or an enforced guard. Prompt delivery does not establish compliance. The single ReAct-style consumer owns action choice, application APIs own effects, and the evaluation driver owns the stated 30-step task limit. SRC-1 §3.3.3, §§4.1.2–4.1.5; RT-BAP-guidelines-channel and RT-RTE-agent-execution.
The representative documented workflow builds subtask memory from earlier train/dev trajectories, selects guidance for a later request, executes the agent loop, and evaluates final application state. It is not a witnessed end-to-end run. Consumer-led checkout correction illustrates recovery, but no trace attributes that correction to retrieved memory. Agent-declared completion, delivery of memory, and benchmark success are different events. AppWorld counts a task successful only if all tests pass and a scenario successful only if every variant passes; that oracle evaluates outcomes, not individual tips. A failed score is not an established automatic repair, rollback or online-write trigger. SRC-1 §2, §§3–4; workflow account and EPI-RTE-benchmark-check.
Reported contribution and independent property assessments
On held-out AppWorld tasks, the paper reports a baseline of 69.6% task goal completion (TGC) and 50.0% scenario goal completion (SGC). Subtask memory with model-guided selection reports 73.2%/64.3%; subtask cosine reports 73.8%/57.1%. These support attributed configuration-level outcome differences, not independent causal validation. Benefit is not universal: task-level cosine at threshold 0.5/top-3 reports 66.7%/48.2%, below baseline, and train easy tasks also decline. Thus the assessment for later benchmark tasks depends on extraction granularity and retrieval configuration, rather than memory delivery alone. Attribution and consolidation were not independently varied, and selector contrasts include filters, ranking and configuration differences. SRC-1 §§4.2–4.4; RT-CLM-reported-results.
The documented automatic trace-to-durable-guidance-to-later-consumer path is a positive design-level trace-learning mechanism. Its learning benefit remains the reported configuration contrast above, not an isolated component effect. Theory-builder conditions remain separate: condition 1, localized explanatory/conditional tips, is claimed; condition 2, content-dependent consumption, is claimed through intended semantic use; condition 3, retained content-directed criticism, is uninspected; condition 4, criticism shaping another round, is uninspected. Tips and fields supply described addressability, not an inspectable assumption editor. No retained challenged-claim/test/blame trace establishes that criticism of a consumed theory improved future capacity. SRC-1 §§3.1–3.3; RT-RTE-trajectory-extraction and RT-RTE-tip-consolidation.
Reflection is claimed through representation of past behavior intended to affect later behavior, with both causal directions unobserved. Computational proposal and successor selection are described, but full operational autonomy and autonomous theory building are not established. Self-improvement has a declared pathway with future task completion as its objective: experience informs guidance, intended retention changes behavior-shaping organization, and prompt delivery supplies an advisory authority path. The fourth link—later operation causally depending on the particular changed guidance—is unestablished. Exercised self-improvement and the claimed recurring feedback cycle therefore remain unverified, independently of the reported gains. SRC-1 §§3–4; RT-RTE-trajectory-extraction, EPI-CLM-feedback-cycle and RT-BAP-guidelines-channel.
Pinned implementation and stored artifacts would change the architectural assessment; delivered-tip/action traces would address activation; independently varied attribution and consolidation would address component effects. Round-linked evidence of criticism, successor admission and content-dependent later action would change the theory-building and exercised self-improvement assessments. No source reread was needed for this synthesis.
Limitations
| Limitation | Affected SRC IDs and record citations | Inspected boundary | Conclusion prevented | Evidence that would resolve it |
|---|---|---|---|---|
| Documented loop, not inspected implementation or execution | SRC-1; RT-RTE-trajectory-extraction, RT-RTE-agent-execution, boundary | Complete frozen HTML, §§3–4; external assets, linked works and deployments excluded | Implemented guarantees, deployed behavior, or witnessed end-to-end success | Pinned pipeline, consumer and evaluator code plus inspectable execution artifacts; bundled figure if visual details matter |
| Outcome and attribution warrant unresolved | SRC-1; EPI-RTE-self-outcome, EPI-RTE-report-interpretation, EPI-RTE-attribution-generation | §§3.1.1–3.1.4 descriptions and examples | Accurate unlabeled outcomes, causal identification, preservation of causal context, or warranted tip transfer | Raw inputs, exact prompts and outputs, independent outcome calibration and causal/transfer tests |
| Successor semantics and maintenance control unresolved | SRC-1; RT-RTE-tip-consolidation, EPI-RTE-conflict-disposition, MEM-OBJ-trajectories | §§3.1–3.2; initialization and physical lifecycle unspecified | Complete admission inventory, faithful merges, tested successor efficacy, actual persistence/reload, withdrawal or rollback | Intake/admission code, before/after corpus and merge decisions, successor tests and versioned lifecycle traces |
| Vector-index relationship unspecified | SRC-1; MEM-OBJ-embeddings, RT-CMP-text-embedding | §§3.2.2–3.3.1; content/purpose and description vectors both described | Complete index schema, compatibility or stale-vector handling | Pinned embedding configuration, index schema, query and reindex implementation |
| Provenance fields do not establish an audit route | SRC-1; MEM-OBJ-tips, MEM-OBJ-metadata, EPI-CLM-provenance-validation | §2.1, §§3.2.3–3.3.3 | Full retained rationale, dereferenceable evidence, or executed effectiveness auditing | Claim-to-source lineage, retained original analyses and an exercised evidence-reading/audit route |
| Selection-unit ambiguity carried from report verification | SRC-1; RT-RTE-cosine-retrieval, memory annotation, RT-CLM-reported-results, verification limit | §3.3.1 says top-k tips; §4.2.2 explains top-3 through matched task descriptions | Exact delivered-tip-count or prompt-size comparison for top-3; no universal five-tip or token bound | Selector implementation and matched-description-to-delivered-tip logs with actual prompt sizes |
| Benefit is reported and treatment-bounded | SRC-1; RT-CLM-reported-results, EPI-RTE-benchmark-check | §§4.1–4.4 aggregate configuration comparisons, including adverse cases | Independently reproduced gains, uncertainty estimates, universal benefit or isolated attribution/consolidation effects | Evaluator code, repeated raw runs, uncertainty reporting and controlled component interventions |
| Delivery does not close the criticism or self-improvement chain | SRC-1; RT-RTE-trajectory-extraction, RT-RTE-tip-consolidation, RT-BAP-guidelines-channel, EPI-CLM-feedback-cycle | §§3–4; initial guidelines and proposed recurring cycle | Exercised reflection, full operational autonomy, conditions 3–4, criticism-induced future-capacity improvement, or exercised self-improvement | Round-linked criticism and revision records, computational role coverage, delivered-guidance/action traces and causal comparisons |
| Provider versions and parameter state opaque | SRC-1; RT-CMP-gpt4, RT-CMP-text-embedding | §4.1.2 family label and §§3.2–3.3 embedding role | Pinned model behavior, endpoint fixity or parameter-learning attribution | Exact model snapshots, provider mutability contracts and any parameter-update evidence |