Signet synthesis

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

Bounded synthesis

Capture evidence, then maintain it

Signet is a shared memory, identity and context-delivery system with shipped host adapters and an internal model-driven maintenance agent. It is not the enclosing foreground agent. This account concerns the repository at commit 12dba1ce753babda735f2ba7b6f150c6c0b7f95e, with an applicability cutoff of 2026-10-14. The functional boundary includes the whole shipped system; inspection establishes selected material routes, not exhaustive implementation coverage. SRC-1 supplies code and target-native doctrine, but no inspected execution trace or causal experiment. External hosts, provider weights, remote services, dependency internals and release binaries are not frozen evidence. See the boundary and evidence and Source register. This synthesis adds no source reads.

The principal progression is host activity → retained evidence → maintained semantic state → selected context for later work. Its control is split: host callbacks and daemon rules own capture, scheduling and delivery; a routed model chooses maintenance tools and proposes content. The daemon admits or rejects the effects. There is no single model conversation joining these responsibilities. See RT-RTE-pi-lifecycle, RT-RTE-dreaming-pass and RT-BAP-maintenance-tools.

In the inspected Pi route, startup sends session, agent, project and runtime identities to daemon hooks. Conflicting runtime claims can be rejected. Terminal session-end handling queues capture and returns zero saved memories plus queue status; that response is not completed semantic ingestion. The later capture worker strips injected memory context, canonicalizes turns, rejects richer or divergent retained-history mismatches, and writes/indexes completed evidence. Failed attempts have failed/dead accounting, not a demonstrated lossless-recovery guarantee. Canonical transcript files, session content and audit references remain distinct retained parts. See RT-RTE-pi-lifecycle, MEM-RTE-transcript-capture and MEM-OBJ-session-transcripts.

Capture is not the only admission path. Direct remember admits caller text through daemon checks without invoking the maintenance model; arbitrary caller identity does not establish human approval of each item. Imported documents retain distinct readable projections, canonical structured material or chunks. Host compaction supplies summaries whose original selection and surviving verbatim tail belong to the external host. These sources have different lineage and cannot inherit one common truth warrant. See RT-RTE-direct-remember, MEM-OBJ-memory-artifacts and MEM-OBJ-session-summaries.

Delegate a bounded maintenance task

Dreaming, Signet's background maintenance process, admits work according to enablement, inference availability, capacity, pressure, triggers and scope leases. Scheduled work can span installation scopes; an explicit user request narrows the grant to the requested agent. A fresh worker receives maintenance guidance, scoped history, attention, clock and supplied tools. The model owns subsequent tool choices. In the Pi executor, resource discovery is disabled and parent-side wrappers reject off-scope agent inputs. These are source-configured capability restrictions, conditional on the Pi dependency honoring them—not OS sandboxing. ACPX and non-Pi provider execution have a different, incompletely inspected resource envelope. Local and hybrid-localhost authentication exceptions also prevent treating token scopes as a universal deployed isolation guarantee. See RT-RTE-dreaming-pass, RT-CLM-tool-boundary, RT-RTE-alternate-executor and the runtime account.

The representative workflow is one incremental-content pass: admit scoped work, inspect evidence, propose semantic changes, stage a complete memory-head entry set, return from the executor, and undergo database finalization. A staging acknowledgment is not durable delivery. A second submission after an accepted stage clears that stage; ending without another accepted stage fails finalization. Late checks can reject a changed base revision or invalid evidence. Earlier ontology batches can already have committed, so pass failure does not imply whole-pass rollback. The prompt's user-review treatment of ambiguous relationships is doctrine, not proof that every write receives human approval. See WF-RTE-head-finalization and the workflow account.

The curated memory head separates natural-language entries from symbolic revision membership and retained support. Authorized, base-matching, bounded requests reaching entry commit must supply exact current scoped evidence for each support. However, unchanged current body content or an already empty head can return a successful no-change result before submitted-support validation; these returns do not replace retained provenance. Carried support is separately revalidated for maintenance reads and clearing guards. Other semantic operations have structural or hygiene provenance alternatives, so neither “every write has an episodic quote” nor “every successful finalization freshly validates support” is justified. See MEM-OBJ-memory-head, EPI-RTE-head-support, EPI-RTE-head-publication and RT-RTE-dreaming-pass.

These checks establish source correspondence and operational admissibility, not truth or entailment of the interpretation. No actual semantic input/output pair establishes whether Dreaming preserves, derives or adds propositions. Publication acceptance is not an evidence-consuming acceptance of factual truth against a substantive criterion. The public source-linking claim therefore has selected implementation support, not an established whole-graph invariant or truth guarantee. See EPI-CLM-source-connections and EPI-RTE-head-support.

Deliver selected state, not the entire record

Database head completion, file projection and model receipt are separate outcomes. Publication can fail after database acceptance and can be retried; user-authored MEMORY.md is preserved rather than overwritten. Startup inspects generated content against authoritative current-head state, withholds unverifiable or stale generated material, then redacts and budgets the eligible text. Explicit identity configuration or disabled recent context can omit this route. The delivered head body omits supporting quotes, although maintenance can inspect retained support. See MEM-OBJ-memory-head and WF-RTE-head-finalization.

Pi startup automatically selects background context and appends hidden messages; explicit Pi/MCP recall instead fulfills a query and returns selected tool text. Startup can trim or suppress already selected material, so candidate retrieval does not guarantee delivery. Recall can display a graph claim with at most two evidence items and further redaction/truncation. These channels supply advisory model input, not permission to execute historical requests or control over the host's tools. See RT-BAP-host-context, RT-RTE-pi-lifecycle, MEM-RTE-explicit-recall and MEM-OBJ-entity-attributes.

Internal maintenance also reloads accumulated context. Selected pass/tool records become model-generated history leaves; adjacent lines become parent summaries. Later scoped passes receive folded history and can request children or a retained pass record through zoom. Eligible non-running passes include failed passes, not just completed successes. Keeping children provides a reload opportunity but does not prove that summaries preserve reasons or advertise every omitted detail. The nominal 512-byte line and 24,000-byte folded-view targets are not hard ceilings. A fresh restricted worker keeps earlier full work outside its initial window, but durable tool writes escape before the worker ends; its only output is not a final return value. See MEM-OBJ-dreaming-history, MEM-RTE-history-compaction and report-verification.

Recall can itself produce durable state. Aggregate recall retrieves evidence, can plan bounded follow-up queries, asks a routed model for a stable note and defaults to saving eligible output. Evidence-class and surface-form gates, deduplication, source links and query hints constrain saving; they do not test entailment. A rejected save can still deliver unsaved synthesis, and embedding failure can leave saved text without a vector. Prior aggregate notes are excluded from aggregate searches to limit recursive synthesis. This is a semantic-write route outside Dreaming, not merely passive retrieval. See MEM-RTE-aggregate-recall, EPI-RTE-aggregate-guard, EPI-RTE-aggregate-retention and EPI-RTE-aggregate-delivery.

Withdrawal changes future eligibility: tombstones, supersession, linked staleness and head-currentness checks can exclude material from inspected current-memory consumers. It does not retract already delivered host context or establish universal erasure, transitive invalidation or restoration. See MEM-RTE-state-withdrawal and EPI-RTE-memory-supersession.

Assess retained change separately from improved performance

The strongest supported learning-related contribution is implemented trace-fed change with named later consumers. Captured sessions can feed durable semantic/head content; pass/tool records feed retained history; host compaction produces retained continuation context; retrieval membership, paths and ratings feed durable selector weights. The last route changes code-owned graph ordering, not provider model weights. Automatic admission and later-consumer wiring support these local trace-learning findings without requiring new knowledge or demonstrated improvement. Improved capacity for future action remains unestablished: no operation or comparison demonstrates that these changes improve decisions or retrieval. See MEM-RTE-transcript-capture, MEM-RTE-history-compaction, MEM-OBJ-session-summaries, MEM-RTE-path-feedback, MEM-BAP-feedback-selection and profile-verification.

The accepted runtime does not establish a formulated-theory route meeting theory-builder conditions 1–4. Separately: localized method guidance is supplied, but its classification as a formulated theory is not established; delivery does not establish decisions depending on that theory's content; content-directed criticism placing blame and changing reliance is not established; retention of history does not establish iteration on such criticism. Criticism of a consumed theory improving future capacity is likewise not established. These are bounded evidential gaps, not whole-system absences. Reflection is also not established: retained accounts of prior work and writable state alone do not prove both causal directions of a selected self-representation. Computational scheduling, proposals and admission establish automatic roles, not an autonomous theory builder or the autonomy of the excluded host's full loop. See RT-RTE-dreaming-pass, RT-OBJ-pass-prompt, MEM-OBJ-dreaming-history and the epistemic bounded conclusion.

For self-improvement, the feedback-to-selector route supplies evidence-responsive durable change capable of affecting Signet's later retrieval organization. It is a concrete partial pathway, not merely an external work product. But rating validity and an independently specified improvement objective are unresolved, and no subsequent operation is shown to exercise the changed path and causally depend on it. Exercised self-improvement and successful improvement are therefore not established over this analysis horizon. Other optimization and adoption routes remain incompletely inspected; this is not a repository-wide negative. See MEM-RTE-path-feedback, MEM-BAP-feedback-selection and report-verification.

Scenario matters. For later-session reference and continuity, Signet implements capture, scoped maintenance and conditional context delivery. For justified factual reliance, those same paths establish provenance and eligibility rather than truth acceptance. For testing answer fit, MemoryBench provides a separate reference/rubric or protocol evaluator, checkpoint retention and resume selection; an incorrect answer can complete evaluation without vetoing product memory. The reported 97.6% accuracy is not a verified result here, nor would answer success alone identify a component's causal effect. See EPI-RTE-benchmark-judgment, EPI-RTE-benchmark-checkpoint and EPI-CLM-benchmark-accuracy.

The assessment would change with inspectable capture-to-maintenance-to-later-consumption traces, checked semantic input/output comparisons, validated evaluators and controlled outcome comparisons. An implemented truth-acceptance criterion with a consequential consumer would change the warrant account. A retained, content-directed criticism that shapes the next round would change the theory-builder account. Independent variation of feedback or a maintenance component, together with later behavior and outcomes, would support finer improvement attribution. Complete adapter, recovery and adoption traces would change the presently partial coverage—not retroactively turn existing wiring into observed benefit.

Limitations

Limitation	Affected SRC IDs and record citations	Inspected boundary	Conclusion prevented	Evidence that would resolve it
Partial material-route inventory; boundary table also omits a top-level plugins row for already-included responsibilities.	SRC-1; MEM-RTE-explicit-recall, MEM-OBJ-memory-artifacts, MEM-OBJ-session-summaries, MEM-OBJ-identity-files, RT-RTE-pi-lifecycle; report-verification.	Selected capture, import, maintenance, recall and Pi routes; provider synchronization, importer reconciliation, non-Pi adapters, hierarchy/thread-head reload, predictor lineage, editor approval and comprehensive recovery remain unresolved.	Complete admission, source-consistency, cross-host delivery, recovery or negative acceptance comparisons.	Route-specific inspection of unresolved producers, admission gates, selectors and consumers; complete plugin coverage mapping.
Incomplete evaluation/improvement plane.	SRC-1; MEM-OBJ-run-checkpoint, EPI-RTE-benchmark-judgment, EPI-CLM-benchmark-accuracy; report-verification.	Benchmark judging/resume inspected; verification also finds metric/failure retention and declared canary patch selection, but not an established automatic production-change loop. Other protocols, calibration and adoption consumers remain uninspected.	Verification of reported accuracy, causal component attribution or repository-wide absence of optimization/autonomous improvement.	Inspectable scored runs, validated references/judges, controlled comparisons and traced release/adoption consumers.
History route's completed-pass wording is too narrow.	SRC-1; MEM-RTE-history-compaction, MEM-OBJ-dreaming-history; report-verification.	Eligible non-running passes with tool-call evidence include failed passes.	A completed-success-only history inventory; observed fidelity or complete failure retention.	Corrected eligibility description and input/summary/reload instances covering failures and exclusions.
Coarse partial profile units bundle distinct objects, consumers and derivation paths.	SRC-1; MEM-OBJ-session-summaries, MEM-OBJ-identity-files, MEM-OBJ-run-checkpoint, MEM-OBJ-path-feedback; profile-verification.	Compaction/recovery, identity and benchmark text/control grouping; host-summary, identity and feedback lineage grouping.	Object-by-object inventory counts or comparisons of which individual origin/consumer is completely classified; positive existence findings remain supported.	Separate classification units with their own coverage warrants in a later analysis.
No observed execution or causal comparison; external implementations and release assets are not frozen.	SRC-1; RT-CMP-inference-router, RT-RTE-alternate-executor, RT-CLM-tool-boundary, RT-BAP-host-context; boundary.	Repository source configuration and selected adapter/daemon wiring, not deployed host/provider/OS behavior.	Exercised delivery, reliable model uptake, deployment isolation, exact model fixity, binary equivalence or benefit.	Registered execution evidence and deployment/dependency evidence at a newly declared boundary; interventions for causal attribution.
Provenance and persistence do not settle semantic warrant; successful no-change head finalization need not freshly validate submitted support.	SRC-1; MEM-OBJ-memory-head, EPI-RTE-head-support, EPI-RTE-head-publication, EPI-RTE-aggregate-synthesis, EPI-RTE-aggregate-guard.	Exact-source checks, no-change branches and aggregate eligibility/style gates.	Universal fresh support validation, truth/entailment, accepted ampliative knowledge or post-acceptance lifecycle integration.	Checked source/output propositions and substantive acceptance criteria with consequential consumers; explicit validation covering no-change requests if that guarantee is intended.
Context reduction and withdrawal have bounded guarantees.	SRC-1; MEM-RTE-history-compaction, MEM-RTE-state-withdrawal, WF-RTE-head-finalization, MEM-OBJ-session-summaries.	History folding/zoom, staged head finalization, current-memory eligibility and startup freshness inspection.	Hard history byte ceilings, lossless summary/reload, pass-wide rollback, universal transitive invalidation/restoration or erasure from existing host windows and copies.	Complete retention/recovery and invalidation traces; enforced budget bounds and semantic-fidelity checks where required.
Retained adaptation does not establish theory-directed learning, reflection or exercised self-improvement.	SRC-1; RT-RTE-dreaming-pass, MEM-OBJ-dreaming-history, MEM-RTE-path-feedback, MEM-BAP-feedback-selection.	Selected maintenance/history and feedback-to-ordering paths; rating origin/validity, formulated criticism and later causal dependence unresolved.	Improved future capacity, theory-builder conditions 1–4 as a complete conjunction, reflective/autonomous theory building or exercised/successful self-improvement.	Localized theory and content-dependent decisions, retained blame-bearing criticism shaping another round, bidirectional self-representation evidence, and objective-relative updates traced into later operation and compared outcomes.