MCMA-OpenMemory synthesis

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

Bounded synthesis

Evidence and architectural boundary

MCMA-OpenMemory is a user-controlled, encrypted memory archive with shipped CLI and web consumers and optional model services. Its Librarian and SecurityAgent are deterministic PHP service roles, not evidence of a model-directed worker scheduler. This account covers the repository at commit 556954d95f0d6d6894c5d46ffcc25a907150b0a7, with an applicability cutoff of 2026-11-08. The whole-system functional boundary includes capture, maintenance, retrieval, model-facing prompts, browser recovery and evaluators. The evidence is source implementation and attributed design, not target execution. There are no supplied observed runs or causal experiments; tests are inspected programs, not passing results. External services and actual deployed configuration remain outside inspection. See SRC-1 in the boundary and Source register.

The central distinction is between retaining content, assigning it reuse status, and consuming it through a particular route. An encrypted revision does not make its assertions true. Nor does one validation policy govern all copies or consumers. The library uses content-addressed encrypted objects and a replaceable manifest/root index to locate current revisions. Previous hashes retain lineage. Canonical explicit memories, their knowledge mirrors, interactions, derived navigation and diagnostic objects share this infrastructure but have separate update and consumption policies. See MEM-OBJ-library-index, MEM-OBJ-explicit-memory, MEM-OBJ-knowledge-metadata and MEM-OBJ-interaction.

From a request to an answer

The representative workflow is a browser question that needs generation, followed by archive-based interruption recovery. It was selected because it connects request identity, retrieval, context construction, generation, capture and delivery in one shipped path. The browser stores a pending request/conversation identity; the server authenticates the principal, selects the user library and checks for an archived request. Explicit-save and mutation syntax intercept before ordinary asking. Otherwise PHP, not the model, owns next-step selection. Billing-enabled requests add provider reservation and settlement protocols; those wrappers do not cover every possible provider caller. See the workflow account and RT-RTE-web-request.

For an ordinary non-broad question, reusable exact knowledge can terminate the request. Otherwise configured semantic retrieval embeds the question, ranks actor-visible current-hash candidates and can return a reusable candidate meeting similarity or configured rerank thresholds. No generation call does not necessarily mean no model call: embedding may already have occurred. A recognized broad-recall question follows a different policy: matching confirmed canonical memory can return before strict knowledge assessment, while recognition suppresses exact and semantic direct returns even if canonical recall finds nothing. That latter forcing case qualifies the analysts' simpler exact-first description, as the report verification records. See RT-RTE-ask-answer and MEM-OBJ-semantic-vectors.

Strict knowledge assessment checks state, confidence, freshness/current requirements and reuse policy. It rejects disputed/retracted and never-direct records; other failures request revalidation. This is an implemented eligibility gate, not an independent investigation of the answer. Provenance count counts supplied entries, not corroborating investigations. Canonical recall instead maps confirmed maturity to verified/0.95 and stale=false without receiving currentRequired. Thus current=true is neither a universal freshness guarantee nor a request for fresh external research. Knowledge-only retraction does not establish withdrawal of the canonical source. The source itself qualifies confidence and provenance as metadata, not proof. See EPI-RTE-knowledge-assessment, EPI-CLM-knowledge-policy, EPI-OBJ-knowledge-record and EPI-RTE-knowledge-validation.

When no direct route returns, the absence of a configured generator yields provider-required. Otherwise code constructs context and calls one generator, rejecting empty text rather than conducting a factual review or model-selected revision cycle. The CLI is narrower: its ask branch does not supply the three web context builders or archive the interaction. See RT-RTE-ask-answer and the workflow account.

What reaches the model

Web context builders reload selected durable material rather than compacting an accumulating live agent window. Conversation selection combines recent anchors with lexical relevance and can include unverified turns while excluding disputed/retracted ones. Multi-memory selection uses assessed knowledge and can deliver stale/current-required material as marked revalidation context. Broad recall has its own subject search and quality policies. Item limits, field clipping and byte accounting leave originals in storage for later reads. These budgets do not bound all combined provider messages, wrappers or exact tokenizer output. There is no isolated subagent frame whose return is the sole escaping content. See RT-RTE-context-supply and its memory annotation.

The shipped connectors serialize selected reference fields and uncertainty into model input and separately instruct the model not to obey instructions inside memory. This is an implemented instruction policy requiring model cooperation, not an enforcement boundary. The provider receives selected data, not a Library handle. Direct custom provider calls need not use the selectors. Delivery is wired; actual reliance, compliance and benefit are unobserved. See RT-BAP-memory-reference.

Capture, correction and later reuse

Generated capture is optional software admission: it defaults to unverified/0.5 and preserves an existing exact record instead of overwriting it merely because generation succeeded. Explicit saving is different. A user supplies save intent and source text; an optional model organizes it; software accepts structurally valid output or falls back to source text. Code then writes confirmed canonical memory and attempts a verified knowledge mirror. Preservation instructions do not constitute a meaning-equivalence check, and the source original is not necessarily the text later returned. See RT-RTE-ask-answer, RT-RTE-explicit-capture and RT-OBJ-organizer-instructions.

Caller validation, answer replacement, interaction approval and canonical mutation provide distinct correction paths. They change later eligibility or content, but owner-role invocation alone does not establish human deliberation or evidence-consuming factual acceptance. Approval can validate existing knowledge without replacing it with the archived answer. Canonical changes, mirrors, indexes and derivatives are not shown to commit atomically; successful canonical writing can coexist with failed synchronization. Tombstones and retraction withdraw current reliance on covered routes, not historical ciphertext. See MEM-RTE-knowledge-editing, RT-RTE-interaction-approval and RT-RTE-memory-mutation.

The strongest supported trace-fed contribution is automatic durable organization: archived turns produce discovery state used by later context lookup; optional catalog classification produces retained labels that select thematic destinations. These are implemented trace-fed updates with later consumers, not merely raw log retention. They do not establish improved capacity for future action. Deterministic thematic snippets supply browsing material and are excluded from canonical recall; context traces supply inspection, not an established automatic policy-revision loop. The independently checked profile verification preserves these distinctions. Supporting records are MEM-OBJ-conversation-index, EPI-RTE-catalog-classification, MEM-RTE-thematic-sync and MEM-RTE-inspector-snapshot.

No supplied evidence establishes that criticism of a consumed theory improved future capacity. The accepted account does not establish a formulated-theory route meeting conditions 1–4, and storage or labels cannot substitute for localized theory content, content-dependent decisions, content-directed criticism or retained iteration. It also does not establish reflection's two causal directions, autonomous theory criticism, or exercised self-improvement. These are independent unestablished properties, not consequences of a negative overall grade. The deterministic ownership of this request workflow is supported; it is not evidence of autonomous improvement. Configured inference requests do not update model weights, but opaque external provider state cannot support a system-complete learning negative. See the epistemic bounded conclusion, RT-RTE-ask-answer and MEM-RTE-semantic-maintenance.

Delivery and scenario-relative assessment

Server JSON return, knowledge capture and conversation archiving are separate outcomes. The web attempts archive/trace writes even with remember=false, and persistence failure need not cancel a successful answer. Browser recovery polls for an archived conversation/request pair without regenerating. Pending means no archive entry was found, not that a computation is running. Recovered answers omit original context and decision fields; best-effort persistence and preflight lookup do not establish exactly-once execution. See RT-RTE-web-request and WF-RTE-request-recovery.

For recalling user-retained material, the implementation supplies revisioned storage, explicit edits, multiple requested retrieval routes and bounded reference reconstruction. For fresh factual answering or audited review, its labels and similarity gates do not supply the missing factual evaluator, uniform freshness policy or complete replay evidence. Actor-aware permission checks can deny covered operations, but trusted PHP code holding raw APIs and keys is outside that protocol's isolation claim. See MEM-RTE-library-authoring, EPI-RTE-knowledge-assessment and RT-CLM-permission-scope.

Concrete changes that would alter these assessments include a shared freshness/withdrawal policy across canonical and knowledge consumers; semantic comparison before promoting organized text; transaction or reconciliation evidence across mirrors and indexes; and recoverable decision traces alongside answers. Supplied executions would establish operation only within their conditions. Demonstrating improved capacity or exercised self-improvement would additionally require tracing an evidence-responsive retained change into later behavior, with comparison evidence appropriate to any claimed benefit or attribution. None is established by the present source wiring alone.

Limitations

Limitation Affected SRC IDs and record citations Inspected boundary Conclusion prevented Evidence that would resolve it
Broad-recall forcing case is underdescribed in the analyst progression; carried from report verification. SRC-1; RT-RTE-ask-answer; report verification AskService guards both exact and semantic returns when broad recall is recognized. Not every reusable exact hit terminates; generation need not imply nonreusable knowledge. The synthesis preserves this exception. Corrected route account covering broad recognition with no qualifying canonical result.
Direct web validation's embedding visibility field is overstated; carried from report verification. SRC-1; EPI-RTE-knowledge-validation, MEM-RTE-semantic-maintenance; report verification Web confirmation calls existing-vector/hash refresh, not embed. No embedding call or content transmission follows from that validation field. Corrected delegation field distinguishing refresh from inference indexing.
No supplied execution or causal comparison. SRC-1; Source register, RT-RTE-ask-answer, RT-BAP-memory-reference Frozen implementation, doctrine and attributed deployment report; tests not run. Actual activation, reliability, compliance, capacity improvement and exercised self-improvement are not demonstrated. Inspectable executions linking retained changes to later consumers; controlled comparisons for benefit and attribution.
Semantic fidelity and factual acceptance remain unresolved. SRC-1; RT-RTE-explicit-capture, EPI-RTE-knowledge-validation, EPI-RTE-catalog-classification Preservation instructions, structural checks and caller-assigned validation; no concrete evidence-consuming reviews supplied. No warranted claim that organized text preserves meaning or that generated assertions were accepted against factual evidence. Concrete input/output and review records naming evidence, criterion, intended use and scope.
Separate object revisions and nonuniform consumer policies. SRC-1; RT-RTE-memory-mutation, RT-RTE-explicit-capture, RT-RTE-ask-answer Canonical commit, mirror/index updates and knowledge assessment are distinct. No atomic coherence, universal current-data guarantee or canonical withdrawal inferred from knowledge-only retraction. Cross-object consistency/recovery evidence and a traced common freshness/withdrawal contract.
Context and derived displays can omit qualifications or retain stale eligibility. SRC-1; RT-RTE-context-supply, MEM-OBJ-thematic-summary, MEM-OBJ-context-catalog Clipped fields, bounded excerpts and hash-based inspector refresh. No complete-prompt token bound, faithful full-context representation or live eligibility inferred from inspector counts. End-to-end payload accounting, fidelity checks and time-aware cache reassessment.
Recovery is best effort and answer-level. SRC-1; WF-RTE-request-recovery, RT-RTE-web-request Browser pointer, archive/index lookup and partial reconstruction. No durable completion, exactly-once execution or full decision replay guarantee. Failure/concurrency traces, durable admission controls and complete reconstruction evidence.
External state and generic payload consumers remain unresolved. SRC-1; RT-CLM-permission-scope, MEM-OBJ-vault, MEM-RTE-library-authoring, MEM-RTE-semantic-maintenance Actor-aware protocol, logical object layout, configurable model endpoints and exposed callback capabilities. No deployed isolation, exhaustive physical-backend/form/lineage classification, exact weight stability or arbitrary model-secret route. Actual configuration, adapter/provider identity evidence and specified consumer/admission contracts.
Learning and evaluator coverage is bounded, not an exhaustive negative. SRC-1; MEM-OBJ-conversation-index, EPI-RTE-catalog-classification, MEM-RTE-inspector-snapshot; report verification Supported discovery/catalog updates; generic caller routes, historical decryption, remote internals and evaluator assertions not exhaustively traced. Positive trace-fed wiring does not enumerate all learning routes; no universal portability, security or evaluator-absence conclusion. Accepted records covering named alternatives and relevant evaluators, with operation/benefit evidence where claimed.