AI-Context-OS synthesis
Type: agentic-system-analyses/types/agentic-system-synthesis.md
Bounded synthesis
A local memory layer, not an enclosing agent controller
AI-Context-OS, branded MEMM, turns a local workspace into retained knowledge and instructions for external agents and its own desktop chat. This account covers the shipped desktop, CLI, MCP transports, generated instructions, state-admission routes and evaluators at public commit 9ed263080cadf98702bebb42d5260b45bd526f6c, with an analysis cutoff of 2026-10-18. SRC-1 supplies implementation and doctrine/design evidence. The repository identifies this as an earlier public snapshot; private continuation and later releases are not assessed. No target, test or provider execution, real user vault, observed run or causal experiment was supplied. Architectural wiring is therefore distinguishable from activation, correct use and benefit. See the boundary and evidence and Source register.
Control is divided. A caller requests context or a write; Rust code selects retained material; a configured provider produces chat text or ingestion candidates; another caller command decides proposal application. The inspected chat consumer returns provider text without a tool-execution scheduler or answer oracle. Desktop HTTP MCP shares workspace state, while CLI and embedded stdio servers initialize their own state. External hosts retain responsibility for task execution and instruction loading. See RT-RTE-desktop-chat, RT-CMP-inference-provider and the runtime account.
From captured material to eligible knowledge
Canonical memories are Markdown files containing an L0 summary, L1 body, optional L2 detail and symbolic control metadata. These are retained authoring surfaces, not evidence of automatically compacted conversations. Prose reaches human/model interpretation; metadata separately controls scoring, dependencies and specified mutation guards. Files hold canonical, staged and operational objects; SQLite holds observability records and maintenance suggestions. Neither graph-based scoring nor JSON encoding establishes a separate graph or vector store. See RT-OBJ-memory, MEM-OBJ-l0, MEM-OBJ-l1, MEM-OBJ-l2, MEM-OBJ-memory-meta and MEM-OBJ-observability-db.
The representative implemented workflow is capture → proposal generation → caller disposition → optional canonical write. Capture stores text, link previews or files in staging. Nontext attachments can remain opaque bytes; normalization changes state or previews, not binary extraction. Proposal generation constructs a heuristic fallback, optionally asks a provider for JSON, and persists a candidate with rationale. Duplicate-discard precedence survives inference replacement; inference errors retain the heuristic rather than retrying the provider. The provider receives selected item fields and schema, not assembled vault context or an independently scheduled worker frame. See MEM-RTE-inbox-capture, MEM-OBJ-inbox-attachment, RT-RTE-ingest-proposal and the workflow account.
A separate apply/reject command owns disposition. Promotion and source routing write canonical content; update_memory and needs_review only change workflow state. Application does not require Pending state or a confidence threshold, so persisted rejection is not an irreversible veto. Successful promotion means a written memory with attempted refresh and a returned Applied proposal, not demonstrated later retrieval or a correct answer. Writes are not transactional: later errors can follow disk mutation, and router regeneration failure need not undo admission. Proposal rationale remains separately retained rather than automatically accompanying canonical query results. See RT-RTE-ingest-proposal, EPI-OBJ-proposal-control, EPI-OBJ-proposal-content and the memory annotations.
For staged ingestion, this provides durable candidate construction and caller-controlled disposition. It does not establish evidence-consuming acceptance against a truth criterion, intended use and scope. JSON shape, hash/URL equality, protection checks and coded conflict terms have real but narrower domains. “Canonical,” “source,” confidence and protection do not warrant factual reliability. Model-generated proposal semantics remain indeterminate without actual input/output propositions. The stronger governance-before-promotion design is explicitly a feature proposal, not an observed guarantee. See EPI-RTE-governance-checks, EPI-CLM-canonical-files and EPI-CLM-governed-promotion.
Selection changes what each consumer receives
Requested context queries rescan files and combine lexical/BM25 signals, graph proximity, importance, recency, access frequency and skill dependencies. Code greedily allocates existing L1/L2 bodies under estimated costs, with at most three L2 allocations. Required dependencies relax a selection condition but reserve no budget; a forced L0-only item can appear in scored output without entering either assembled package branch. MCP returns rule summaries, selected bodies and unloaded L0 references. Chat omits that prelude and unloaded list. Simulation exposes scored metadata, not an inventory of delivered bodies. See RT-RTE-context-query, RT-OBJ-context-result and the memory annotations.
For a query-driven context scenario, the implemented contribution is selective external-content loading, not whole-window management. Estimates do not charge all prelude, reference, header or history text. MCP skill loading independently returns the requested skill and immediate dependency L1 bodies without this allocator; it does not recursively load a full dependency closure. Static adapters afford host-directed file reading, while connector handoff exports IDs and L0 summaries rather than full allocated content. Requested retrieval is pull; unresolved consumers prevent a system-wide pull-only assertion. See RT-RTE-skill-load, RT-RTE-adapter-handoff and the profile verification.
Desktop chat rebuilds vault selection for each send but accumulates prior turns without inspected trimming or compaction. Disabling vault context clears only the dedicated context string and IDs: history and the caller's system prompt can still contain vault information. Enabled requests with nonblank caller context bypass reconstruction. These are channel controls, not provenance filtering or confidentiality isolation. Selected text advises the provider; successful transport establishes neither obedience nor answer truth. See RT-RTE-desktop-chat and RT-BAP-context-provider.
Admission and protection are similarly route-specific. Root inbox exclusion does not cover every nested inbox, and direct canonical saves bypass proposal review. Folder-contract warnings do not universally reject parseable files. Structured/raw/MCP mutation checks differ; folder validation does not validate every independently supplied filename input. Journal save can add canonical date tags without the structured protected guard, and restore overlays files outside that guard. These findings prevent universal protection or confinement claims without establishing a deployment exploit. See RT-RTE-workspace-write, MEM-OBJ-memory-meta, MEM-RTE-journal-derive and MEM-RTE-backup-restore.
Feedback has operational reach, not demonstrated improvement
The strongest supported adaptive contribution is a wired trace-fed loop: selected MCP loads, skill/required-dependency reads and desktop reads automatically update durable access counts and timestamps; later scans feed them into ranking and decay. A second wired path converts retained serving telemetry into persisted maintenance advice for later maintenance readers. Both qualify as trace-fed behavior-shaping writes at this architectural level. Their inputs record exposure events, not judged task success. Neither requires semantic rewriting to exist, but changed rankings, improved future capacity and causal benefit remain unobserved. See MEM-RTE-access-feedback, MEM-OBJ-memory-usage, MEM-OBJ-observability-db, RT-RTE-optimization-status and the profile verification.
Maintenance has several distinct consequences. Human governance actions can delete content or increase importance; “archive” is deletion, not preserved historical withdrawal. Optimization apply only changes suggestion status, not canonical content or policy. Health scores compute telemetry-based proxies, not factual accuracy. Session summaries are stored separately from the daily log used by consolidation, with no established later session-log reader in the inspected search. Thus the evidence supports priority adaptation and advice, not a unified autonomous episodic consolidation pipeline. See MEM-RTE-governance-actions, EPI-RTE-importance-boost, EPI-RTE-health-score, MEM-OBJ-session-log and MEM-OBJ-daily-log.
These contributions do not establish the independent stronger properties. The accepted records do not establish theory-builder conditions 1–4 as a complete route: retained prose alone cannot establish a formulated theory, delivery cannot establish content-dependent consumption, and diagnostics or scores do not establish formulated content-directed criticism and retained iterative revision. Nor is improved capacity through criticism of a consumed theory established. Telemetry and later priority dependence supply parts relevant to a reflective account, but the records do not establish a complete reflective theory-building mechanism. Automatic derivative writes are supported; autonomy of the caller-disposition workflow is not, because the provider does not own admission. No system-complete negative follows from these local findings. See the epistemic bounded conclusion, RT-RTE-ingest-proposal and the report verification limits.
Exercised self-improvement is likewise unestablished at this public-system boundary. The access loop reaches future selection inputs, but no independently assessed improvement objective or subsequent operation causally dependent on the change is supplied. Maintenance advice has an even earlier gap: marking it applied does not implement the proposed organizational change. This is not a denial of the wired trace-fed contribution. A linked before/after execution showing retained changes entering later selection or inference would establish activation; a controlled outcome comparison would be needed for benefit or component attribution. Evidence-consuming admission records, actual generated propositions and later reason use would change the epistemic assessment. Implementing a maintenance executor would change the status-only account, but would still not by itself prove improvement. See MEM-RTE-access-feedback, RT-RTE-optimization-status and EPI-RTE-source-tests.
Limitations
| limitation | affected SRC IDs and record citations | inspected boundary | conclusion prevented | evidence that would resolve it |
|---|---|---|---|---|
| Partial mutation/recovery inventory and security assessment; carried from report verification | SRC-1; RT-RTE-workspace-write, MEM-RTE-onboarding-write, MEM-RTE-backup-restore; verification limit | Guarded APIs, onboarding control and overlay restore; full templates, every rename/move/duplicate path, symlink/deployment permissions, cross-vault isolation and live-database consistency remain unsettled | Universal path confinement, protected-file immutability, exact rollback, consistent live-database recovery or cross-vault isolation comparisons | Complete material mutation/permission traces, template and vault-switch audits, and consistent snapshot/restore evidence under stated deployment conditions |
| Epistemic lens is not a complete semantic inventory; carried from report verification | SRC-1; MEM-RTE-journal-derive, MEM-RTE-onboarding-write, MEM-RTE-backup-restore, epistemic conclusion; verification limit | Detailed epistemic routes plus operational records for journal/task/template/restore, not a full semantic assessment of those branches | System-complete negative about semantic production, acceptance or theory-building capacity | Route-specific semantic and acceptance accounts for these remaining operations, with content, criteria and consequential consumers |
| Implementation evidence only; no supplied operation or experiment | SRC-1; RT-RTE-desktop-chat, MEM-RTE-access-feedback, RT-RTE-ingest-proposal, EPI-RTE-source-tests | Frozen source, evaluator code and callable workflow | Exercised human review, correct provider use, empirical relevance, exercised self-improvement, improved capacity or causal benefit | Linked ingestion/retrieval/inference traces and outcome evidence; controlled intervention comparisons for causal attribution |
| Public snapshot and opaque provider/host boundary | SRC-1; source register, RT-CMP-inference-provider, RT-RTE-adapter-handoff | Earlier public revision; provider interface and generated host artifacts | Current-private-product conclusions, exact model identity, provider-internal memory/learning classification or universal host compliance | Separately frozen newer evidence, immutable model identity and inspectable provider/host retention and consumption contracts |
| Missing provenance, opaque payload semantics and incomplete later consumers | SRC-1; MEM-OBJ-inbox-attachment, MEM-OBJ-session-log, MEM-RTE-onboarding-write, RT-RTE-adapter-handoff, EPI-OBJ-proposal-content | Known local writes and readable projections; arbitrary original vault contents, attachment meaning, generated proposition relations and later session/scratch consumers unresolved | Complete lineage, authority, read-back or trace-source inventories; guaranteed extraction, semantic synthesis or unified episodic consolidation | Original-content provenance, actual model input/output, extraction implementations and named later consumers with consumption evidence |
No target-source files were read for this synthesis; all findings use the supplied members. Reconciliation reports no unresolved conflict. Profile verification declares no additional limits; its partial and not-determinable classifications do not become complete inventories here.