IsaacCLupus_mnemosyn_spec synthesis
Type: agentic-system-analyses/types/agentic-system-synthesis.md
Bounded synthesis
From sources to proposed knowledge
Mnemosyne / Jarvis Knowledge OS is a proposed shared memory substrate for external agents, not an enclosing agent controller. This account covers the specification, implementation studies, SQL schema and superseded Python bridge at commit 0fda14eb33d2fc01a4e3969a9ab487b22c4f9d92, with an applicability cutoff of 2026-10-15. The boundary is complete artifact, partial loop: a coherent specification is available, but its unified runtime is not established. SRC-1 supplies doctrine/design, static implementation and attributed prototype operation, not observed runs or causal experiments. External backends and live clients are outside the frozen source set. No target code was executed. See the boundary and evidence and Source register.
The proposed progression is raw-source ingestion → fast-model extraction → heavy-model article compilation → audit → approval or rejection → publication → later query. The core delegates extraction, compilation, audit and finalization to a Python satellite through named JSON files. Finalization prepares a result; publication belongs to the proposed Rust core. These are file-level process contracts, not demonstrated independent agents completing a task. The workflow account follows this source-connected progression; RT-RTE-design-lifecycle records its runtime ownership and limits.
Compilation is intended to retain source provenance and generate linked, cited articles. Rendered prompts include source text, related-page summaries and flagged contradictions. Their length truncation is not demonstrated token accounting, and the illustrated model call places the whole rendering in one user message rather than enforcing the recommended instruction/source separation. Neither a typed article nor a citation request establishes semantic fidelity or truth. Without concrete source/output pairs, the generated content's relation to its premises remains indeterminate, not established derivation or ampliative conjecture. See RT-OBJ-compile-template, RT-BAP-compile-guidance and EPI-RTE-article-generation.
Checking, admission and rejection
The design distinguishes structural lint, contradiction detection and adversarial review from the subsequent publish/reject decision. Structure checks concern conformance; agreement with existing pages does not certify those pages or settle which conflicting claim is wrong. A score threshold and a human veto, when configured, are intended to control wiki admission. Configured automatic approval is another policy alternative, so unconditional human approval is not a system-wide guarantee. No concrete evidence-consuming acceptance of a scoped new claim is supplied. See EPI-RTE-article-audit and EPI-RTE-article-disposition.
The strongest supported learning-related contribution is the proposed retained-criticism path: rejection feedback persists separately from a rejection counter, then supplies rewrite instructions to a later ArticleWriter. The scheduler sketch uses rejection and hand-edit state to constrain retries. A second proposed path extracts facts and memory candidates from logged conversations, uses heat and decay for promotion, and admits committed memories through human approval. These are design-level trace-fed durable updates with intended later consumers, not merely log retention. Direct profile visibility remains undecided. See MEM-RTE-feedback-recompile, MEM-BAP-feedback-guidance, MEM-RTE-conversation-promotion and MEM-OBJ-user-profiles.
What remains unresolved is whether such criticism actually improves future capacity. Feedback plumbing has schema/scope defects, no exercised rewrite is supplied, and no comparison attributes a capacity gain to criticism. Heat selects relevance rather than truth. The profile verification supports the local claimed trace-learning classifications without converting them into parameter training, successful improvement or complete memory coverage.
What the legacy bridge actually wires
The shipped jarvis-kb driver loads configuration and dispatches initialization, publication, proxy serving, MCP serving or status. An external client's MCP request selects a fixed handler. Ingest, compile and audit delegate to external Synto, Synthadoc and Link tools; they are not automatically chained into the proposed lifecycle. Compilation does not call publication. Audit results return as advisory diagnostics without a local gate consuming them to block publication or withdraw pages. See RT-RTE-foundry-tools and EPI-RTE-audit-dispatch.
Separate publication copies Synto and Synthadoc Markdown into a unified vault, normalizes metadata, requests indexing and snapshot/load operations, and optionally commits files to Git. Bodies survive, but approval, hash and other governance fields do not all survive metadata replacement. Numeric confidence takes precedence over contradicted status: a source with confidence 0.9 and contradicted status receives an established maturity label. This is deterministic labeling, not factual evaluation. Synto draft directories are excluded; the Synthadoc traversal has no equivalent filter, and neither copier checks per-page approval or write guards. See RT-RTE-vault-publish, MEM-OBJ-unified-vault and EPI-RTE-maturity-mapping.
Files are written before external commands whose nonzero exits can be ignored. Printed completion therefore does not establish synchronized recipients; optional Git history does not provide coordinated rollback. Source deletion is not reconciled with copied destinations. This bridge is best-effort publication, not the proposed atomic promotion/index/graph/pack transaction or authenticated human-edit protection. See RT-CLM-publish-atomicity and RT-CLM-write-guards.
Direct remember is another admission path: the gateway forwards caller content to Link without implementing the proposed inbox/audit/approval sequence locally. An HTTP exception triggers CLI fallback, which omits the HTTP path's project field. Failure after a remote write can leave an ambiguous outcome before fallback; no local deduplication or compensating transaction is established. Link owns the actual admission and retained state. See RT-RTE-link-remember.
Delivery and later work
A later ask or search requests retained backend material. The gateway returns backend output as MCP text; it does not own the client's next step. Ask echoes a context budget without forwarding it, history budget is unused, and search's source label does not restrict backend requests. Backend retrieval, synthesis and history are opaque. The proxy separately forwards model requests: queue mode prioritizes waiting chat work and serializes one instance's calls without preempting a running job; router mode and direct backend calls bypass that queue. Configured model aliases do not establish immutable versions or parameter updates. See RT-RTE-client-query, RT-RTE-model-proxy and RT-CMP-heavy-model.
Proposed graph and pack routes offer different later-consumer mechanisms. Graph retrieval is seeded by page identity; automatic model-input assembly supplies primary text and up to five outbound relationship summaries, not all retrieved neighbor bodies. Positive depth does not enforce a one-hop traversal, and schema/content-field mismatches prevent treating the sketch as executable wiring. Exported packs afford later file-aware agent consumption, but missing-file entries, filename collisions and reused-directory residue prevent complete snapshot or withdrawal guarantees. Requested retrieval and automatic context insertion are distinct operations; neither establishes unsolicited memory delivery to an external foreground agent. See MEM-RTE-graph-context, MEM-RTE-pack-export and MEM-BAP-query-knowledge.
Independent properties and scenario-relative assessment
For a scenario requiring a specified source-to-query workflow, the artifact supplies explicit handoffs, differentiated checks and feedback-bearing rejection. For a scenario requiring enforced approval, atomic delivery, bounded model context or reliable retraction, the inspected bridge does not establish those guarantees. Future claim/verdict/dissent governance and optional integrity proofs remain separate proposals: verdict fields do not demonstrate acceptance, and a Merkle anchor would concern serialized-byte integrity rather than claim truth. See EPI-RTE-claim-governance and EPI-RTE-proof-export.
The independent assessments in RT-RTE-design-lifecycle remain at their own evidence strengths:
- Theory-builder condition 1, localized content: claimed. Templates solicit claims/explanations, but no resulting theory is inspected.
- Condition 2, consumption: claimed. Related pages and contradictions are intended inputs; dependence on a particular theory is unobserved.
- Condition 3, content-directed criticism: uninspected. Contradiction reporting does not establish placement of blame and resulting revision or changed reliance.
- Condition 4, iteration: claimed. Retained feedback is intended to shape another round; exercised theory-directed iteration is unobserved.
- Criticism improving future capacity: uninspected. The proposed update paths are the positive contribution; no retained-result comparison establishes this stronger learning conclusion.
- Reflection and autonomy: uninspected. Content audit does not establish two-way self-representation of operating machinery. Full computational ownership of the internal theory-building roles is not established; human gating is configurable, not uniformly absent.
- Self-improvement: not established as exercised at this boundary. Retained pages could affect later organization through query or compilation, so an article label alone does not exclude self-change. Here, however, proposed updates have not been shown entering a live authority path and causing subsequent operation to depend on the change. Neither membership in exercised self-improvement nor successful improvement follows.
A wired unified runtime with enforced admission, coordinated rollback and verified consumer-side context assembly would change the implementation assessment. Linked source/candidate/audit/decision/rewrite/read-back traces would change the operation assessment. A comparison isolating retained criticism's contribution to later capacity would be needed for an improvement attribution. No new source reads were made for this synthesis.
Limitations
| Limitation | Affected SRC IDs and record citations | Inspected boundary | Conclusion prevented | Evidence that would resolve it |
|---|---|---|---|---|
| Unified runtime and external backend effects are not observed | SRC-1; RT-RTE-design-lifecycle, RT-RTE-client-query, RT-RTE-link-remember | Specification, static bridge and interfaces; no target execution or backend internals | Successful admission, retrieval, persistence, deployed privacy/isolation and end-to-end benefit | Unified implementation and backend contracts, plus linked execution and consumer traces |
| Context-removal inventory omits a separate satellite helper | SRC-1; report verification limit, MEM-RTE-graph-context, memory Read-back | Verification inspected python_thoughts/04-llm-client-glue.md: helper drops oldest non-system messages while preserving system messages and at least two non-system messages |
Complete comparison of proposed context-removal mechanisms or guaranteed budget fitting; approximate token counting and retention floor can prevent fitting, and wiring/reload are unestablished | Separate helper route and caller/reload account, tokenizer and retained-floor analysis, consumer-side context evidence |
| Evaluation/admission inventory omits an ArticleWriter confidence alternative | SRC-1; report verification limit, MEM-RTE-feedback-recompile, EPI-RTE-article-disposition, workflow recovery | Watcher study uses citation-marker count, wikilink count and length; diagram allows confidence-threshold auto-approval | Complete inventory of proposed admission criteria or inference that every confidence score comes from model review; heuristic is not a factual evaluator | Separate heuristic-to-threshold route assessment, activation policy and evidence of downstream decisions |
| Publication lacks coordinated admission, recovery and withdrawal | SRC-1; RT-RTE-vault-publish, MEM-OBJ-unified-vault | Complete legacy copier and its external command calls | Universal approval/write protection, atomic synchronization or reliable retraction; original backend authorship and exact Link transfer remain unresolved | Enforced per-path admission, metadata preservation, transaction/recovery and deletion contracts, with traces |
| Retained-feedback wiring and retry ownership remain incomplete | SRC-1; MEM-RTE-feedback-recompile, workflow annotation | Selected watcher planner/writer and complete compilation study; incompatible DDL references and inconsistent lookup | Executable feedback loop, complete retry/stopping policy or criticism-induced improvement | Compatible state schema, resolved selector and retry policy, linked rewrite/later-use evidence and comparison |
| Profile visibility and verification are unsettled | SRC-1; MEM-OBJ-user-profiles, MEM-RTE-conversation-promotion | Canonical fields and draft promotion proposal | Direct profile injection, complete profile admission/privacy policy or warranted extracted facts | Defined profile consumer, selection/privacy/admission controls and fact-verification procedure |
| Graph and pack studies do not establish complete delivery | SRC-1; MEM-RTE-graph-context, MEM-RTE-pack-export | Inspected study selection/assembly/export snippets | Executable graph context, guaranteed token/depth bound, complete pack snapshots or revocation | Compatible schema/content access, bounded traversal and token assembly; collision-safe fresh exports and withdrawal protocol |
| Concrete knowledge-production warrant is missing | SRC-1; EPI-RTE-article-generation, EPI-RTE-article-audit, EPI-RTE-markdown-normalization, EPI-RTE-claim-governance, EPI-RTE-proof-export | Proposed generation, checks, URL-to-title normalization, draft governance and proof export | Specific warranted derivation or accepted ampliative claim, reference fidelity, executed verdict/revocation or truth from integrity proofs | Source/output pairs, checked reference mapping, scoped evidence-consuming decisions and actual withdrawal consumers |
| Memory alternatives and study coverage remain partial | SRC-1; MEM-OBJ-state-metadata, RT-RTE-client-query, RT-RTE-model-proxy, profile verification | Recorded opaque backend/history/job/provider state, optional embedding/cache paths and selected studies | Complete-value memory statistics, exhaustive context/admission inventory, fixed model internals or system-wide absence claims | Operative payloads, provenance, admission and consumer routes for each unresolved unit; broader relevant study inspection |
| Learning, theory building and self-change lack stronger evidence | SRC-1; RT-RTE-design-lifecycle, MEM-RTE-feedback-recompile | Proposed criticism/update routes without concrete theory episode, live self-change or intervention | Condition 3, exercised theory iteration, improved future capacity, reflection, autonomous theory building or exercised/successful self-improvement | Concrete blame/revision/iteration records, role ownership and self-representation evidence, later causal dependence on retained changes, and outcome comparisons |