From Agent Memory to Portable Skills synthesis
Type: agentic-system-analyses/types/agentic-system-synthesis.md
Bounded synthesis
From workspace experience to a reusable instruction product
NAMS's strongest supported contribution is a described route from retained agent experience to durable, reviewable instructions for later agents, with separately maintained evidence links. The target is its skill-distillation and governance subsystem: a builder or improvement plane, not the enclosing agent runtime or all of NAMS. Evidence is the Neo4j article frozen as neo4j-article-2026-10-18-99f91d4da959, applicable through 2026-10-18. SRC-1 supplies design descriptions and attributed operation, but no implementation, generated skill, review decision or execution trace. Thus the account establishes a documented architecture, not verified operation or benefit. Linked documentation, remote media and the AIP evaluation are outside the capture (boundary, Source register).
The representative workflow is a scoped request becoming a published skill available to another agent. A caller submits scope through HTTP, MCP or a dashboard; a Redis-stream job hands progression to the distiller. It resolves conversations, takes a workspace snapshot lock, waits for extraction to settle, consolidates recurring actions by frequency and recorded outcome, derives topology, requests textual synthesis, applies gates and persists a pending product in Neo4j. This is construction of instructions, not execution of their task. The initiating caller is not necessarily human, and initiating construction does not approve its generated content (RT-RTE-skill-distillation; workflow account).
The backing material includes conversation and reasoning text, recorded tool calls and statuses, and ontology entities. These parts have different roles. Text supplies synthesis evidence; fixed-role call fields support outcome separation and bindings; graph identities support reference checks and later grounding repair. Arbitrary tool payloads are not necessarily symbolic merely because they are graph properties. Original authorship, prior transformation, extraction fidelity and status-update provenance remain unresolved (MEM-OBJ-memory-text, MEM-OBJ-tool-call, MEM-OBJ-memory-graph).
Narrow model delegation and bounded admission
The architecture separates deterministic topology from model-written claims and annotations. The procedure graph holds typed bindings, dependencies and branch structure derived from recorded traces. The natural-language package holds the procedure, claims, names, descriptions, completion criteria and references. These share Neo4j storage without sharing execution semantics. Fewer than three recorded steps yields prose rather than a skeletal graph; not every skill is therefore a typed graph (RT-OBJ-procedure-graph, RT-OBJ-skill-package, RT-RTE-skill-distillation).
A Python synthesis endpoint receives at most 400 interleaved, PII-redacted snippets. The model proposes claims and annotations, not topology. This bounds model-visible input, but does not establish complete visibility: snippet ranking, preferential omission and reload of omitted evidence are unspecified. The separate call is a narrow synthesis handoff, not evidence of a general sub-agent stack or host conversation compaction. Model version and parameter lifecycle are independently uninspected (RT-CMP-llm, RT-RTE-skill-distillation).
The synthesis service drops claims lacking sources; the worker checks reference existence against the subgraph. These controls concern provenance presence and identity, not source truth or entailment. Subsequent gates check grounding, recurring-pattern coverage, coherence, specification and PII. Grounding must clear 0.9 and coverage 0.6, but their formulas and error behavior are unavailable. A scope containing separable procedures is withheld with a split suggestion, not automatically rewritten into successful jobs. Passing permits pending persistence, not acceptance as warranted general knowledge (EPI-RTE-source-check, EPI-RTE-quality-check, EPI-RTE-pending-admission).
These are declared ordered controls, not demonstrated invariants. The lock concerns concurrency, not sandboxing or tool permissions. Effective locking, settled extraction, queue delivery, graph transactions and synthesis-service integrity are external contracts. Initial gates cannot be extended to prove equal enforcement on the prose fallback, direct writers or repaired patches. Unresolved route reachability establishes neither bypass exclusion nor a demonstrated bypass (RT-CLM-admission-controls).
Publication and later consumption
Persistence hands a pending skill to a human reviewer, who reads steps and supporting evidence and approves or rejects publication. This is a distinct admission decision after software construction. No substantive review rubric or actual acceptance decision is supplied. Evidence links created before review establish evidence-linked retention, not post-acceptance integration of a warranted claim (RT-RTE-review-delivery, EPI-RTE-pending-admission).
A published SKILL.md package can be loaded into Claude Code, fetched through MCP or consumed over REST. Its downstream force is interpreted procedural guidance and reference context. The host owns task selection, tool access and execution; package delivery grants neither permissions nor mandatory graph compliance. Requested evidence retrieval and client fetching support documented pull routes, while automatic host selection and maintenance triggers remain unresolved. The workflow's deliverable is an available reviewed package, not demonstrated receipt or task success (RT-BAP-skill-consumption, RT-RTE-review-delivery).
Maintenance without a guaranteed recovery loop
Governance compares a step's expected tool status with the current grounded ToolCall status. A mismatch can flag contradiction even while every citation remains. This is a meaningful check beyond reference existence, but its domain is status consistency: it does not diagnose whether a failure came from the procedure, tool, data or policy. Flag consumption by publication or host execution is unspecified (RT-RTE-drift-check).
For a drifting step, repair takes the workspace lock, follows merged sources to survivors, drops deleted sources, revalidates the graph and creates a superseding patch, leaving other steps untouched. With no surviving grounding it refuses and flags full re-distillation. This supports described local grounding revision, not guaranteed restoration of tool success. Automatic drift-to-repair invocation, mandatory patch re-review, installed-copy refresh and rollback are unresolved; these adjacent routes cannot be joined into an observed automatic recovery chain (RT-RTE-step-repair).
What the contribution does and does not establish
Automatic trace-fed construction of durable guidance supports a claimed trace-learning mechanism: conversational records, tool traces and linked decision/action/result experience feed an update whose product has documented later consumers. Human publication review does not erase that automatic construction. Improved capacity for future action remains uninspected, however; neither persistence nor instruction delivery demonstrates that capacity increased. No evidence shows criticism of a consumed explanatory theory improving later capacity. The article's reported structured-versus-prose AIP comparison concerns another evaluation and cannot identify a NAMS distillation, gate or governance effect (RT-RTE-skill-distillation, RT-BAP-skill-consumption, EPI-CLM-graph-validity).
The theory-builder conditions remain separate unresolved findings: (1) no actual whole explanatory theory unit is supplied; (2) no decision trace shows dependence on its content; (3) no formulated criticism assigns blame and changes a theory or reliance; (4) no retained criticism is shown shaping a next theory round. Step addressability and source-survivor repair do not establish these conditions, but opaque processing does not establish their absence either (RT-RTE-step-repair).
Reflection is independently unresolved: backing-memory/status representation is described, but not a representation-mediated loop concerning the builder's own method. Internal construction is computational in design; initial publication explicitly requires human judgment, so the documented release workflow is not fully autonomous. An autonomous theory builder is not established. Self-improvement of this bounded builder is also unestablished: its maintained output guides external agents, and no evidence traces a change entering the builder's own live organization and shaping subsequent operation. These are not a single negative grade, nor do they negate its described instruction-maintenance contribution (RT-RTE-review-delivery, RT-RTE-step-repair, RT-BAP-skill-consumption).
For reconstructing a reusable procedure from scoped work, the design offers inspectable provenance, separate proposal/admission roles and bounded local repair. For establishing that the procedure is justified, transfers safely or improves performance, the supplied evidence is insufficient. Concrete input/output specimens and review criteria could resolve semantic transformation and acceptance; implementation and failure traces could establish control coverage; consumer traces and controlled comparisons could establish activation and benefit. A builder self-update trace would be required separately for self-improvement attribution (EPI-CLM-actionable-knowledge, RT-CLM-admission-controls). No source reread or additional evidence was used for this synthesis.
Limitations
| Limitation | Affected SRC IDs and record citations | Inspected boundary | Conclusion prevented | Evidence that would resolve it |
|---|---|---|---|---|
| Document-only evidence; linked resources excluded | SRC-1; boundary, RT-RTE-skill-distillation | Frozen article body and captions; subsystem only | Implemented guarantees, observed successful workflow, whole-NAMS coverage or independent interoperability | Pinned implementation, job/export/review artifacts and receiving-host traces |
| Backing-content provenance and payload forms unresolved | SRC-1; MEM-OBJ-memory-text, MEM-OBJ-tool-call, MEM-OBJ-memory-graph | Described distillation inputs, not upstream ingestion implementation | Complete lineage/form inventory, historical accuracy or trusted current-status provenance | Stored input examples, extraction/recording code and source histories |
| Selection and semantic transformation opaque | SRC-1; RT-RTE-skill-distillation, EPI-OBJ-claims, RT-OBJ-skill-package | Bounded snippets, deterministic topology description and graph/prose alternatives | Complete context, preserved explanations, checked entailment or demonstrated novel claims | Selector rules, actual scoped inputs, outputs and exported references, including thin scopes |
| Gate and release coverage not verified | SRC-1; RT-CLM-admission-controls, EPI-RTE-quality-check, RT-RTE-review-delivery | Named initial triggers, gates and human review; alternate writers and patch release unspecified | Universal review enforcement, privacy/correctness assurance, path equivalence or bypass claims | Gate definitions and implementation, fallback tests, reader/writer policies and patch publication rules |
| Queue, locking and recovery contracts unspecified | SRC-1; RT-RTE-skill-distillation, RT-RTE-step-repair | Described Redis job, workspace lock and Neo4j writes | Exactly-once admission, crash safety, production isolation or completed retries | Transaction, retry, idempotency and lock lifecycle implementation with failure traces |
| Maintenance does not close the downstream effect chain | SRC-1; RT-RTE-drift-check, RT-RTE-step-repair, RT-BAP-skill-consumption | Status flags, grounding patches and described package delivery | Causal diagnosis, restored success, automatic withdrawal/refresh, complete pull-only inventory or known push selector | Scheduling, flag/version consumers, patch review and revocation code plus host traces |
| Semantic acceptance, learning benefit and theory properties unestablished | SRC-1; EPI-CLM-actionable-knowledge, RT-RTE-step-repair, RT-CMP-llm | No actual theory, review decision, model lifecycle or content-dependent consumer trace supplied | Warranted general knowledge, improved future capacity, theory-builder conditions, reflection or exercised builder self-improvement | Actual claims and evidence-consuming review criteria; criticism/revision/iteration and later-use traces; separately pinned model lifecycle and builder self-update evidence |
| Performance contrast is secondhand and outside NAMS | SRC-1; EPI-CLM-graph-validity | Article's attributed AIP evaluation, without underlying artifacts | NAMS benefit or isolated component attribution | NAMS-specific controlled comparisons with treatment, baseline, outcomes and confound analysis |
| Partial authority inventory omits separate ontology-content use | SRC-1; MEM-OBJ-memory-graph, RT-RTE-skill-distillation, profile-verification limit | Accepted partial inventory classifies ontology identity validation/repair but not its separate evidence role at distillation | Complete part-to-consumer authority comparisons, especially ontology-content use versus identity validation; existing knowledge findings remain supported | Explicit classification of the recorded ontology-content evidence consumer, alongside identity-checking and repair consumers |