ai-memex-cli synthesis
Type: agentic-system-analyses/types/agentic-system-synthesis.md
Bounded synthesis
From captured material to durable knowledge
ai-memex-cli is a Markdown knowledge system whose deterministic CLI supplies tools and delegates semantic maintenance to external coding agents. It aims to capture sources, integrate them into a cross-linked wiki and make accumulated knowledge available to later sessions. This account covers the shipped repository at commit 4adb274a33cbb9ff0f03a6aeeb9c8d2525a5ac3f, with applicability cutoff 2026-10-15. SRC-1 supplies inspected implementation and separately attributed doctrine/design. The whole-system repository boundary includes commands, prompts, maintenance and later-consumer integrations, but not external agent inference, provider persistence or live services. No real vault, execution trace or causal comparison was supplied. Wiring therefore does not establish successful operation or benefit (boundary, Source register).
Capture has distinct admission paths. Direct fetch writes extracted web content with URL/date attribution; keyword capture can require an operator's source selection or admit all results automatically. Delegated capture asks an external agent to retrieve and write material. These operations acquire claims rather than verify them. Extraction can omit source context, and a repeated direct fetch can replace a same-named raw file. Thus the vault's description of raw sources as immutable is a maintenance policy, not a system-wide storage invariant (RT-RTE-capture-source, EPI-RTE-fetch-render, MEM-OBJ-raw).
Session capture also has two materially different meanings. The no-argument default mechanically converts recognized session messages to Markdown; it does not extract lessons. Structured conversion can preserve nonempty system text and some textual tool messages while discarding nested tool results and arguments. If nothing parses, it renders the entire input as user text. The retained document is neither a guaranteed complete transcript nor a verified redacted record. Delegated distillation instead requests decisions, practices and rationale, plus a brief learning note appended to the live vault schema. Those are requested model outputs, not supplied artifacts (RT-RTE-session-distill, MEM-OBJ-session-markdown, EPI-RTE-session-render).
Delegate, write, inspect and try again
The representative workflow is memex ingest <target> --loop, because one shipped controller actually joins delegation, inspection, another attempt and delivery. The caller supplies a target hint, vault and optional agent/cap. Each child receives freshly loaded schema and catalog, the target and any prior lint findings. It runs in the vault directory and discovers sources through its own tools. The external model selects semantic edits and admits them directly into live wiki files, with requested index/log updates. The CLI owns dispatch; it does not validate an approved edit set or enclose edits in a rollback transaction (RT-RTE-ingest-write, workflow account).
This separates maintenance context from the foreground conversation: memex passes selected inputs rather than replaying the entire parent session. Filesystem effects and child output cross back. Temporary prompt cleanup is not evidence that the provider discards child history or that an enclosing host evicts context. Raw preservation and discussion before writing remain interpreted instructions, not a filesystem sandbox or enforced human veto. Starting a command is not approval of each resulting claim (RT-OBJ-ingest-prompt, runtime control cases).
After ingestion, mechanical lint reloads the wiki and checks orphan pages, link-target membership and frontmatter. These checks produce meaningful but narrow structural findings. They do not check factual bodies, citation support or source reliability; provenance fields are not all required. Even linking a page from the root catalog need not satisfy the wiki-to-wiki orphan predicate (EPI-RTE-mechanical-check, MEM-OBJ-page-metadata).
Nonzero findings can become up to 200 repair priorities in the next child's prompt. Zero issues permits a clean stop and best-effort context refresh. The criterion accepts structural completion, not factual content. Ingestion failure can still be followed by clean lint on unchanged files; cap or stall leaves prior edits in place. The selected ingest loop defaults to three iterations, whereas watch defaults to unlimited, and force can bypass the repeat guard. No convergence, bounded spend or rollback follows from this feedback mechanism (RT-RTE-watch-repair, EPI-RTE-loop-disposition, workflow recovery).
Semantic criticism is not absent. A separate optional semantic-lint child is asked to inspect contradictions and other content defects, apply safe repairs and retain ambiguous findings for review. However, a findings page is ordinary retrievable wiki prose, not a quarantine or an implemented mandatory adjudication queue. No supplied finding establishes correct criticism, completed review or evidence-consuming factual acceptance. Repository-local instructions also afford a foreground scan and user review, so child delegation is not uniform across all shipped instruction variants (RT-RTE-semantic-lint, MEM-OBJ-lint-results, report-verification limit).
Make retained material available to later work
The retained store combines natural-language bodies and guidance with symbolic access fields. Page bodies supply context to interpreting agents; IDs, names, tags and descriptions feed fixed ranking or link predicates; schema instructions guide maintenance; registry entries route host-file refresh. These are different consumers and forces, not one blanket authority granted to everything stored in the wiki (MEM-OBJ-wiki-body, MEM-OBJ-page-metadata, MEM-OBJ-wiki-schema, profile verification).
Later use can pull search results and full pages, request injection through includes/keywords, or read a copied project subset. Search caps and filters can omit relevant material, and backend errors can appear as empty results. Injection accepts task/token options without using them to select or budget content. Keyword selection ranks metadata, not page bodies. These routes establish delivery mechanisms, not complete retrieval or accurate answering (RT-RTE-context-delivery).
Automatic supply also exists: management calls preselect schema/catalog for a child, and clean-loop refresh writes a bootstrap into registered host files. The latter contains consultation instructions and, in full mode, an ordered digest of up to five pages per scene with shortened descriptions, not full evidence. Bound scene names select manifest guidance but do not filter the full digest. Refresh is best effort, and a later external host must load and use the file. Bulk refresh can lose saved scene bindings, while registry entries lack per-vault identity. Forced local copies can leave old unselected files behind; neither refresh nor uninstall revokes material already loaded into a host (RT-BAP-host-bootstrap, MEM-OBJ-host-bootstrap, MEM-OBJ-context-registry, MEM-OBJ-local-projection).
For a source-integration scenario, this architecture supports durable external memory, structural repair feedback and selected later context. The README says contradictions have already been flagged, not adjudicated or resolved; this synthesis withdraws its earlier attribution of resolution to that claim. Failure to demonstrate semantic success does not establish that the README promises resolution. The separate README claim that synthesis reflects everything read remains unestablished by the supplied evidence. Synthetic page labels and citation requests do not identify whether an actual output reshapes, derives or adds a claim. Retention and operational use do not establish accepted new knowledge or integration following acceptance. Foreground answer saving is itself instruction-mediated: the installed query reference requires a save request, while vault doctrine asks the agent to file durable answers (EPI-CLM-memory-compounds, EPI-CLM-synthetic-pages, MEM-RTE-query-save).
Adaptation without an improvement result
The strongest supported learning contribution is the wired session-to-durable-guidance pathway: delegated distillation requests a schema learning note, and later maintenance reloads that schema. A parallel summary-to-later-context path can preserve session-derived decisions for reuse. The verified memory classification recognizes these automatic trace-fed write/read routes without requiring observed benefit. The unresolved question is whether useful notes were written, retained their reasons, affected later decisions and improved capacity; no before/after comparison was performed (RT-RTE-session-distill, RT-BAP-schema-prompt, MEM-OBJ-wiki-schema, profile verification).
The independent theory-builder conditions are not established by the mechanical feedback account: 1, no concrete formulated explanatory theory is identified; 2, dependence of decisions on such a theory's content is not shown; 3, localized defects and requested semantic scans do not demonstrate stated theory-directed criticism and changed reliance; 4, forwarding defects demonstrates iteration, but not retention and reuse of the result of theory criticism. These are limits of this account, not claims that opaque model processing lacks those properties. Criticism of a consumed theory improving future capacity is separately unestablished (RT-RTE-watch-repair, epistemic conclusion).
Reflection is likewise unestablished: session capture and schema adaptation provide relevant surfaces, but no supplied account closes both causal directions through a self-representation of the selected system aspects. Autonomy is narrower than success: software can dispatch watcher maintenance without per-edit human approval, while reliable semantic work and autonomy of every theory-building role remain unestablished. Neither uncertainty cancels the known dispatch mechanism (RT-RTE-watch-repair, RT-RTE-session-distill).
For self-improvement at this repository-plus-vault boundary across sessions, schema adaptation is a candidate route into the system's own behavior-guiding organization, not merely a wiki product change. Yet no supplied note or trace establishes objective-bearing evidence determining an actual update, actual organizational change, or subsequent operation causally depending on it. The reload channel alone supports neither exercised self-improvement nor successful improvement. Package update is instead explicit replacement from upstream, not an evaluated self-directed improvement loop (RT-BAP-schema-prompt, RT-RTE-install-update).
A source/output pair with provenance would sharpen transformation and rationale judgments. A linked trace from session evidence through schema edit to a later decision would establish operation beyond wiring. A controlled comparison of later work with and without that retained change would address benefit, with attribution limited to what was varied. Enforced approval, transactional recovery, implemented task/token selection or persistent vault-scoped context bindings would change the corresponding architectural assessment; none can be credited at this revision merely because a prompt requests related behavior.
Limitations
| Limitation | Affected SRC IDs and record citations | Inspected boundary | Conclusion prevented | Evidence that would resolve it |
|---|---|---|---|---|
| Source-attribution caveat: the README claims contradictions are flagged, not resolved | SRC-1; EPI-CLM-memory-compounds | README wording retained in the accepted claim record; no supplied contradiction-handling output | Attributing adjudication or resolution to the README's flagging claim; lack of demonstrated semantic success does not establish such a promise. The distinct all-read synthesis claim and actual semantic success remain unestablished | Explicit source wording would be needed to attribute a resolution promise; linked source/output and review evidence would separately address actual flagging, resolution and synthesis coverage |
| No supplied operational or comparative evidence | SRC-1; RT-RTE-ingest-write, RT-RTE-watch-repair, RT-RTE-context-delivery | Frozen implementation/doctrine; no real vault, child transcript or host-consumption trace | Actual compliance, successful integration, convergence, semantic accuracy, activation and causal benefit | Linked input/edit/output and later-consumption traces; appropriately controlled outcome comparisons |
| Instruction-variant inventory remains incomplete; local lint affords foreground scan, safe fixes, risky-edit user review, findings retention and mechanical recheck | SRC-1; report-verification limit, RT-CLM-management-use, RT-RTE-semantic-lint, MEM-OBJ-lint-results, epistemic boundary | Installed templates and selected repository-local prompts, including the verifier's local lint inspection | Uniform delegation, review or report-path comparison across shipped variants; local doctrine does not prove execution | Complete material instruction-variant mapping and host precedence/activation evidence |
| Source lineage is lossy and warrant is unresolved | SRC-1; EPI-RTE-fetch-render, EPI-RTE-session-render, MEM-OBJ-session-markdown | Extraction and recognized parser/render branches, not every provider format or real input | Full evidential preservation, authenticity, privacy filtering or complete transcript fidelity | Original/captured pairs across supported formats, extraction checks and source-quality evidence |
| Structural cleanliness is not factual acceptance; semantic outputs and review are unavailable | SRC-1; EPI-RTE-mechanical-check, EPI-RTE-loop-disposition, RT-RTE-semantic-lint, EPI-CLM-synthetic-pages | Selected structural predicates and delegated semantic prompts | Semantic consistency, checked derivation, accepted ampliative knowledge or mandatory closure of findings | Concrete propositions, supporting evidence, criticism and evidence-consuming acceptance/review decisions |
| Human approval and protocol precedence are not uniformly enforced | SRC-1; RT-RTE-ingest-write, MEM-RTE-query-save, MEM-OBJ-lint-results | Headless maintenance and conflicting collaboration/save/deletion guidance | Universal per-content approval, raw preservation, read-only query or safe destructive-edit guarantee | Enforced admission controls and traces showing applicable host precedence and actual approval |
| Selection and withdrawal are incomplete across derived copies and host projections | SRC-1; RT-RTE-context-delivery, MEM-OBJ-context-registry, MEM-OBJ-local-projection, MEM-OBJ-host-bootstrap | Include/keyword selection, capped search, bulk refresh, local copy and uninstall paths | Task/token guarantees, persistent scene binding, vault-isolated refresh, complete propagation or revocation | Implemented selection budgets and binding persistence; refresh/withdrawal traces covering downstream consumers |
| Open writes and initial content control remain partially classified | SRC-1; MEM-OBJ-scene-manifest, MEM-OBJ-wiki-body, MEM-OBJ-index-log, profile verification | Known readers/transformations and open model/caller admission, without actual corpus or manifest creation route | Exhaustive lineage, admission, curation, behavioral-effect and trace-origin inventories | Recorded creation/admission provenance and representative input/output/consumer chains for unresolved branches |
| External runtime and fine-grained compatibility remain outside the established account | SRC-1; RT-CMP-coding-agent, RT-OBJ-ingest-prompt, RT-RTE-context-delivery | Caller adapters and subprocess controls; provider/qmd internals and live permission grants uninspected | Pinned model behavior, parameter-update conclusions, host eviction, sandbox guarantees or comprehensive format/security/compatibility claims | Versioned external contracts and scoped permission, context-lifecycle and compatibility evidence |
| Durable adaptation is wired, but theory criticism, reflection and exercised self-improvement are not established | SRC-1; RT-RTE-session-distill, RT-BAP-schema-prompt, RT-RTE-watch-repair, epistemic conclusion | Requested schema additions, future reload and structural-feedback controller | Theory-builder conditions individually, improved future capacity, bidirectional reflective operation, full theory-building autonomy or exercised self-improvement | Retained theories/criticisms and next-round dependence; linked evidence/update/later-behavior traces; separate improvement comparison |