MemexOS synthesis
Type: agentic-system-analyses/types/agentic-system-synthesis.md
Bounded synthesis
MemexOS is a knowledge-production and retrieval system, not an enclosing answering agent. It collects sources, delegates wiki writing, conditionally checks and publishes the changes, and returns retained material to later callers. This account covers the shipped CLI, scheduled maintenance, console actions, prompts, vault instructions and evaluators at commit bb9e9b2ac0adbca87d13a9f7600867128b5f71b9, with an applicability cutoff of 2026-11-07. Its sole source, SRC-1, is the frozen liushengyu/memexos repository. Evidence is implementation text and attributed doctrine, not target execution. No production trace, adopted performance result or causal experiment establishes successful operation or benefit. External model services, qmd internals and ordinary consuming agents' reasoning remain outside the inspected implementation (boundary, Source register).
From collected experience to a selected draft
The system's persistent center is a filesystem and Git vault. Raw sources and derived wiki pages have different lifecycles. Captured text can enter directly, while unattended intake asks a model to admit, reject or retain uncertain candidates. Selected text scanning can veto admission; it is not a guarantee over every binary copy, capture or plugin route. Session intake is already a transformation: selected conversational and tool records become distilled snapshots before wiki compilation. Thus a file in raw/ need not be a verbatim original or independently warranted evidence (RT-RTE-source-admission, MEM-OBJ-raw, MEM-RTE-session-distillation).
The representative workflow is memex ingest, chosen because one shipped caller connects selection, delegation, checking, repair and delivery. Code selects unprocessed, unheld raw paths, obtains a write lock and creates a staging branch/worktree from the vault base. It supplies the writer with source paths, classification and instructions to read the vault rules. The writer proposes wiki, index, log and inbox edits; code retains scheduling and merge authority. No routine per-edit human approval is required. Separate writer and reviewer calls exchange files and selected reports, not an inspected shared conversation (RT-RTE-compile-publish, RT-OBJ-stage, workflow account).
Context selection is consequential at each handoff. Session rendering abbreviates tool output and removes some material; classification uses bounded digests and selected precedents; review receives a capped diff with file-reload access; repair receives rendered problems rather than the entire reviewer session. These mechanisms isolate work and restrict delivered context. They do not establish that every relevant fact reaches the next model. Files can remain available for requested reads without being present in its window. External CLI compaction remains opaque (MEM-RTE-session-distillation, MEM-RTE-catalog-reuse, RT-RTE-backend-execution).
What publication accepts—and what it does not
Gate 1 checks structural properties of the staged changes, including allowed paths, selected link/source references and new-page structure. Optional sentence self-check supplies support diagnostics for selected unmarked additions; unavailable checking can proceed to review. Gate 2 asks a model for source-faithfulness, coverage, overwrite and injection judgments. Deterministic rules then decide publication eligibility over that model-produced inventory. Critical unsupported or contradicted claims and missed critical points veto; noncritical bad claims block above 5%, and must-point coverage blocks below 80%. These thresholds are not independently measured error rates over all actual claims (EPI-RTE-gate1-check, EPI-RTE-selfcheck-check, EPI-RTE-review-claims, EPI-RTE-review-decision).
The acceptance criterion is eligibility of this changed batch for source-faithful vault publication, not truth of the underlying source. The rubric explicitly excludes source trustworthiness. Honestly marked speculation is allowed; PARTIAL judgments and some noncritical errors need not block. Review fields enforce continuation on enabled covered paths, while their reasons guide another writer. On rejection, selected claim, source, location and reason fields reach bounded rework. Exhausted repairs can hold a branch and suppress ordinary retry; review unavailability instead discards the draft and starts a cooldown. Initial writer failure, holding and merge-abort have different recovery paths (RT-BAP-review-verdicts, MEM-OBJ-review, RT-RTE-compile-publish, workflow account).
This is a conditional publication protocol, not the public claim that nothing is published unreviewed. Gate 2 can be configured off; no review-relevant changes can pass without a model call. Catalog synchronization independently publishes material source-use metadata after Gate 1 alone. Stage.publish has no independent review-certificate requirement. Nor is writer isolation uniform: Codex's workspace grant is broader than the writer's named edit paths, and Gate 1 checks the diff rather than undoing prior process effects. The OpenAI-compatible adapter also has a provider-response-shaped fetch dispatch case without a local web-permission recheck; that is code-level reachability, not an observed model bypass (RT-CLM-checked-publication, EPI-RTE-catalog-sync, RT-RTE-backend-execution).
Delivery means the draft merged into the live vault. Persistence adds no epistemic warrant, and no separate post-acceptance semantic integration judgment is established by this compile path. FTS refresh is attempted and vector reindexing follows asynchronously; merge success does not guarantee fresh indexes or a multi-file reader snapshot (EPI-RTE-wiki-retention, RT-RTE-compile-publish).
Later use and feedback
An asking agent requests search, then optionally gets page text. Search combines available lexical/vector routes, filtering, fusion and optional reranking; it can return raw material independently of wiki approval. Get reloads requested file contents. Retained pages supply information, not an enforced answer policy: the caller owns subsequent reasoning, citation and reliance. Classification and source pointers help interpret evidence but do not certify it (RT-RTE-search-get, RT-BAP-retrieved-pages, MEM-OBJ-wiki).
Read-back also occurs inside production. Code supplies pending sources to fresh compilers, URL/author-matched catalog precedents to classifiers, and retained review problems or usage summaries to repair/lint calls. Those automatic supplies differ from requested search/get. Retained state has several forces: prose supplies context and guidance; processing ledgers exclude or select work; verdict fields can veto publication; text and vectors influence retrieval ranking. FTS tokens, dense vectors and external model parameters are distinct objects, not interchangeable kinds of memory (MEM-RTE-catalog-reuse, MEM-OBJ-ledgers, MEM-OBJ-fts, MEM-OBJ-vectors).
The strongest supported adaptation contribution is a wired trace-to-durable-change chain. Session distillation produces reusable text. Usage extraction and model diagnosis can produce feedback; confirmation with nonempty problems creates stored repair tasks; fix can publish revised wiki for later retrieval. A smaller loop persists rerank timing and uses it to set later candidate budgets. These establish automatic trace-fed writes with later consumers, not just passive log storage. They do not demonstrate improved future capacity, accurate diagnosis, better answers or faster retrieval (MEM-RTE-session-distillation, MEM-RTE-usage-repair, MEM-RTE-rerank-budget, profile verification).
The public evolution claim has separate, bounded outcomes. Improvement-directed self-change is implemented dispositionally: evidence can shape repairs, publication changes retained organization, and later writers/retrieval can consume it. Exercised self-improvement remains unestablished because no subsequent operation is shown to depend causally on that change; successful improvement needs additional outcome evidence. Reflection has a partial wiring witness in usage representations that can drive repairs, not an observed faithful two-way cycle. Several internal operations are automated, but whole-system autonomy is not established. Theory-builder conditions 1 (localized formulated theory), 2 (decisions dependent on its content), 3 (content-directed criticism and changed reliance/revision) and 4 (retained criticism shaping another round) remain individually unestablished here: no concrete theory instance is individuated. Neither repair wiring nor schema proposals establish these conditions, and no improvement attributable to criticism of a consumed theory is demonstrated (EPI-CLM-evolves-itself, RT-CLM-schema-change).
Assessment relative to the work
For accumulating sources into a revisable, searchable wiki, the implementation supplies a coherent unattended workflow with explicit rejection and recovery states. For certifying every delivered claim as true, its source-relative rubric, model-selected inventory and alternate publication routes do not supply that guarantee. For demonstrated self-improvement, it supplies the update pathway but not exercised causal dependence or benefit. These are different questions, not a product ranking (RT-RTE-compile-publish, EPI-CLM-reviewed-memory, EPI-CLM-evolves-itself).
The shipped evaluators could supply local evidence about retrieval ranks, answer grading, seeded reviewer errors and tool procedure. Their references and consequences differ: a model-maintained question bank is not independent source truth, and procedure matching can include stubbed effects. No adopted result establishes general quality or an automatic evaluation-driven optimizer. Paired real inputs/outputs would resolve compilation's semantic transformation; longitudinal traces with controlled contrasts could establish whether particular retained changes improve later work. Uniform review claims would additionally require covering the alternate publishers and their actual consumers, not merely renaming a pass result (EPI-RTE-evaluation-bank, EPI-RTE-answer-grading, EPI-RTE-review-calibration, EPI-RTE-tooluse-scoring, EPI-RTE-writer-transform).
Limitations
| Limitation | Affected SRC IDs and record citations | Inspected boundary | Conclusion prevented | Evidence that would resolve it |
|---|---|---|---|---|
| No adopted operational or causal evidence | SRC-1; RT-RTE-compile-publish, EPI-CLM-evolves-itself | Frozen code and doctrine; no production vault or longitudinal trace | Successful compliance, exercised self-improvement, improved capacity and component benefit | Provenance-linked production traces through later consumption, plus appropriate outcome comparisons and independent variation for component effects |
| Acquisition inventory and verification criteria are incomplete | SRC-1; RT-RTE-inbox-maintenance, MEM-OBJ-feedback, EPI-RTE-feedback-check; report verification limit | Ordinary usage verification is traced; inbox local investigation also admits a worker's nonempty, scanned, page/claim-targeted local record as raw feedback before correction | Complete acquisition-route or uniform verification-criterion comparisons; the supported automatic repair chain remains valid | Separate accepted records tracing local investigation's acquisition, admission, delivery and correction consumers |
| Evaluation read-back is locally omitted from the direction inventory | SRC-1; EPI-OBJ-questions, EPI-RTE-evaluation-bank, EPI-RTE-answer-grading; profile verification limit | Retained references reach L1/L2, including a separate grading call, but are not separately classified for direction | Complete evaluation-consumer direction and push-selector comparisons; not the supported pull/push existence findings | Route-level classification of actual reference selection and delivery to each later evaluator |
| Media, research, taxonomy and localized instructions remain incompletely inspected | SRC-1; MEM-OBJ-assets, MEM-OBJ-taxonomy, RT-RTE-inbox-maintenance, MEM-RTE-session-distillation | Selected text and lifecycle routes, not all lower-level media transformations, research payloads, registry decisions or translated prompts | Complete memory inventories, all-route safety/context guarantees, source-warrant preservation and translation equivalence | Accepted records for the unresolved transformations, payloads, admission controls and consumers; direct translation comparison |
| External execution and retrieval internals are opaque | SRC-1; RT-RTE-backend-execution, RT-RTE-extension-loading, MEM-OBJ-vectors | Local adapters, host loader and qmd interface; external provider/CLI/qmd internals excluded | Uniform sandbox enforcement, complete compaction behavior, immutable model identity, complete vector lineage or immediate withdrawal guarantees | Version-specific external contracts and inspectable execution/index evidence at those boundaries |
| Concrete semantic outputs and independent evaluation references are not established | SRC-1; EPI-RTE-writer-transform, EPI-OBJ-questions, EPI-RTE-retrieval-scoring, EPI-RTE-tooluse-scoring | Requested model transformations and inspected evaluator code; generation/qrel provenance and shipped results not fully assessed | An individuated accepted ampliative claim, warranted subject-matter derivation, independent benchmark truth, general retrieval quality or real side-effect success from procedure scores | Paired source/output/acceptance artifacts, independently checked references, complete trace conditions and nonstubbed outcome evidence |
Source-read log: none. This synthesis uses the accepted members and their retained passages; no direct source ambiguity required inspection.