Memex synthesis
Type: agentic-system-analyses/types/agentic-system-synthesis.md
Bounded synthesis
Memex places an editable markdown wiki between supplied sources and later questions. Its daemon retains and schedules work; fresh external Claude Code processes interpret sources, choose files and edit or answer. The strongest supported contribution is this persistent update-and-reuse architecture, including a requested path from filing experience to retained conventions. It is not evidence that accumulated content becomes more correct or that future decisions improve.
This account covers the whole shipped system at commit 1e223b6045be31b70b1ab9e4b28dd403150d29e7, with an applicability cutoff of 2026-10-14. SRC-1 is the frozen https://github.com/wastedcode/memex repository. Implementation and doctrine/design support the account; no populated wiki, execution trace, evaluation result or causal experiment was supplied. Claude Code, provider internals, dependency behavior and configured MCP services are outside the inspected implementation. The boundary and source register establish those distinctions. Both report verification and profile verification have no remaining blockers or declared limits; this does not remove the underlying evidence qualifications.
From supplied files to persistent editorial state
The representative workflow begins with memex ingest <wiki> <files...> into an existing wiki. The CLI uploads local files before submitting a durable job containing their stored names. The daemon checks kernel-derived caller ownership, writes archive bytes and serializes queued jobs per wiki within that daemon. Different wikis can run concurrently. Uploads and external filesystem edits are not covered by that serialization. The runner rebuilds the task, resolves credentials/model/tools and delegates to a fresh Claude process with appended wiki instructions. The caller may be human or automation; selecting sources does not approve each later edit. See RT-RTE-ingest-job, RT-OBJ-queue-jobs and the workflow account.
Claude is asked to read sources, conventions and index, find related pages, extract information, revise or create pages, maintain reciprocal links and reconsider filing conventions. These are steps inside one delegated task, not daemon-enforced editorial checkpoints. Pre-approved Write/Edit tools admit direct filesystem effects without a staged patch or per-edit approval. A consequential layout mismatch limits literal execution: the child starts at /workspace, while scaffolded content is under /workspace/wiki/; prompts name raw/, _schema.md and _index.md relative to the starting directory. Model discovery could compensate, but no trace demonstrates it. Thus the source establishes the delegation and requested procedure, not successful integration into the intended files (RT-RTE-ingest-job).
The retained objects have different jobs. Uploaded bytes are source material, not endorsed assertions. Derived pages supply factual context; _index.md supplies semantic navigation; _schema.md supplies filing conventions; _log.md supplies model-written activity history. Their primary substrate is files, whereas queued tasks, results and coarse daemon audits live in SQLite. Page references and preserved quotations are requested provenance, not a checked per-claim derivation graph. No supplied page demonstrates faithful extraction or complete reasons for later reuse. See MEM-OBJ-raw, MEM-OBJ-wiki-pages, MEM-OBJ-index, MEM-OBJ-schema, MEM-OBJ-log and MEM-OBJ-audit-log.
Delivery, subsequent questions and maintenance
A completed ingest means the child exited zero and its result was persisted for polling clients. Output parsing can fall back to text; neither valid operation JSON nor verified edits determine success. Files can survive even a failed invocation. Reasons may survive in returned job output, but the SQLite audit records coarse lifecycle details, contrary to the documentation's stronger reason-audit claim. Prior completed results are not automatically fetched into later model tasks (EPI-RTE-result-display, MEM-CLM-reason-audit, RT-OBJ-queue-jobs).
A later question is a separate job, not an automatic continuation of ingest. Claude is instructed to read the index, search with grep/glob, read relevant pages and synthesize a cited, wiki-grounded answer. The daemon does not score relevance or assemble selected page contents. API history is supplied verbatim by the caller; the CLI supplies no history. Each job starts without a conversation-resume identifier. Memex therefore carries files, task text and selected configuration across invocations rather than restoring its own full model conversation. Turn/time limits bound work, not token usage or retrieval completeness. Internal Claude compaction remains uninspected (RT-RTE-query-job, RT-RTE-job-recovery).
Wiki-only grounding is a model instruction, not an evidence gate. Query retains Write/Edit access and a requested activity-log append; optional web tools and MCP can enlarge available inputs. The prompt suggests saving a new insight, but does not implement the README's stronger automatic answer-filing workflow. Direct client file browsing is a separate wired read route without model synthesis (EPI-CLM-grounded-query, RT-RTE-query-job, MEM-RTE-file-browse).
Manual lint or ten exit-success ingests counted in current in-memory queue state can schedule maintenance. Restart resets that count. Automatic lint receives a fresh task and access to retained files, not the preceding ingest's operation report; it is not a prerequisite for ingest delivery. Claude is asked to diagnose consistency, possible staleness, links, index and schema drift. Confident structural fixes are intended to be direct; contradictions, merges and stale facts are intended to be report-only. Identical write capabilities do not enforce that partition. Reports persist for readers without an approval/resumption protocol or automatic later-model consumption (RT-RTE-lint-job, workflow annotations).
These routes supply opportunities for criticism, not an independent truth oracle. Source conflict comparison, model confidence and textual conformity have distinct domains. Upload admission, resolve-or-flag disposition, file retention and exit success do not establish evidence-consuming acceptance of a novel factual claim. Without concrete model inputs and outputs, particular semantic transformations remain indeterminate: they could preserve, derive or conjecture content. No warranted-premise derivation, accepted ampliative claim or subsequent integration after such acceptance is established (epistemic bounded conclusion, RT-RTE-ingest-job, RT-RTE-lint-job).
What can change future behavior
The most specific adaptation opportunity is _schema.md: encountered filing patterns are requested to revise durable categories, templates, vocabulary and heuristics, and later ingests are requested to read them. This can change the agent's filing guidance rather than only its factual work product. Activity summaries also have a requested later maintenance consumer. These support claim-layer experience-to-guidance/context paths, not observed learning or a wired historical-log-to-rule transformer. Original trace input classes remain unresolved; the generated log cannot classify its own original inputs (MEM-OBJ-schema, MEM-BAP-schema-guidance, MEM-OBJ-log, profile verification).
The runtime assesses the four theory-builder conditions separately for requested schema maintenance: (1) localized content: claimed; (2) decisions depending on that content: claimed; (3) content-directed criticism and resulting revision: claimed; (4) retained criticism shaping another round: claimed. Generated conventions, actual dependence, explicit blame assignment and next-round effects remain uninspected. File persistence affords a cross-session horizon but does not demonstrate these properties. Whether criticism of a consumed theory improved capacity for future action is uninspected. Reflection is claimed through requested self-comparison, while both realized causal directions remain uninspected. Autonomous theory-building execution is independently uninspected, despite model-owned edit selection. Exercised self-improvement is also uninspected: the schema offers a possible self-change pathway for future filing, but no particular evidence-responsive update, installation and subsequent causal dependence is supplied. No successful-improvement or component-effect claim follows (RT-RTE-ingest-job, RT-RTE-lint-job, epistemic bounded conclusion).
Other behavior changes follow configuration: the runner reloads retained model/tool choices and passes an existing MCP configuration path. Custom .claude.md discovery belongs to external Claude behavior, not explicit text injection by this runner. These fixed configuration/replay supplies differ from requested wiki file reads. Model selection is not weight training, and the default alias does not pin an exact model version (RT-RTE-config-change, RT-CMP-claude-model).
For source-to-wiki accumulation, Memex supplies persistent editorial machinery but leaves fidelity and successful path resolution unobserved. For later answering, it supplies model-directed retrieval but not verified grounding. For interruption-sensitive work, restart replay restores pending eligibility without reverting partial file effects or ensuring exactly-once edits. For confinement-sensitive work, a mount namespace does not hide inherited host paths; the read-only raw alias does not protect all aliases. Whole-wiki withdrawal is destructive even with the shipped keep-data option because DELETE bodies are not parsed; it also does not cancel active children. These are scenario-relative implementation findings, not a product ranking (RT-RTE-job-recovery, RT-CLM-workspace-isolation, MEM-RTE-wiki-withdrawal).
Path-aligned invocations with inspectable file/tool traces would change the activation assessment. Source-to-page-to-answer comparisons with independent grounding checks would change the fidelity assessment. Traced schema criticism, a retained change and later decisions causally depending on it would change the adaptation assessment; measured outcomes and an appropriate comparison would separately address benefit. Transactional or idempotent edit handling, tested confinement boundaries and a repaired DELETE-body path would change the relevant recovery, security and withdrawal findings. The documented sequential prompt-evaluation procedure supplies possible checks, not their results (MEM-RTE-prompt-evaluation).
Limitations
| Limitation | Affected SRC IDs and record citations | Inspected boundary | Conclusion prevented | Evidence that would resolve it |
|---|---|---|---|---|
| No observed operation or causal comparison | SRC-1; RT-RTE-ingest-job, RT-RTE-query-job, RT-RTE-lint-job | Frozen shipped implementation and prompts; no populated wiki or execution results | Successful integration, actual grounding, maintenance reliability and improvement cannot be claimed | Inspectable runs with retained inputs/outputs; independent evaluation and controlled comparisons for benefit |
| Relative-path mismatch | SRC-1; RT-RTE-ingest-job, MEM-BAP-schema-guidance | Child starts at /workspace; scaffolded files are under wiki/ |
Literal intended-file access and schema activation are not established | Trace of compensating navigation, or corrected paths with execution evidence |
| Warrant and semantic transformation remain unresolved | SRC-1; MEM-OBJ-wiki-pages, EPI-CLM-compounding-wiki, RT-RTE-query-job | Requested extraction, contradiction handling and answer synthesis | Faithfulness, novel derivation/conjecture classification, accepted new knowledge and monotonic enrichment cannot be established | Concrete source/page/answer edges and evidence-consuming acceptance decisions with criterion, intended use and scope |
| Adaptation has request-layer support only | SRC-1; MEM-OBJ-schema, MEM-OBJ-log, RT-RTE-ingest-job | Requested schema revision, activity summaries and later reads; no generated conventions or consumed trace inputs | Exercised theory building, realized reflection/autonomy, improved future capacity, exercised self-improvement and original trace-source classification remain unestablished | Retained rules/criticism with blame assignment, original update inputs and later behavior dependent on the revision; separate outcome evidence |
| External consumers and auxiliary state are opaque | SRC-1; RT-RTE-config-change, RT-CMP-claude-model, RT-RTE-job-recovery | Memex invocation arguments and local configuration; external Claude/provider/MCP internals excluded | Complete instruction activation, context/checkpoint behavior, parameter mutability and external memory inventory cannot be inferred | Pinned external implementations/configuration and inspectable consumption/state traces |
| Checks and reports do not enforce semantic acceptance | SRC-1; RT-RTE-lint-job, EPI-RTE-result-display, MEM-CLM-reason-audit | Pre-approved writes, report-only policy, exit-code completion and coarse audit | Verified correctness, enforced human-review stops and automatic reason-to-later-model lineage cannot be claimed | Enforced effect gates and recorded decisions; explicit reason retention/delivery to later consumers |
| Recovery and confinement have narrower guarantees than stronger product readings | SRC-1; RT-RTE-job-recovery, RT-CLM-workspace-isolation | Replay without filesystem rollback; inherited host mounts; external tool enforcement | Exactly-once editing, clean recovery, workspace-only visibility and universal raw immutability are unsupported | Idempotence/transaction evidence, crash traces, enforced filesystem isolation and deployment-specific tests |
| Withdrawal loses preservation choice | SRC-1; MEM-RTE-wiki-withdrawal | DELETE transport defaults keepData false; deletion precedes registry cascades; no child cancellation | The shipped keep-data option cannot be treated as preservation, and deletion is not verified universal revocation | Repaired transport and preservation tests; coordinated cancellation/revocation evidence |
| Input formats, caller control and budgets remain partly unresolved | SRC-1; MEM-OBJ-raw, RT-RTE-config-change, RT-RTE-query-job | Arbitrary uploads/API callers; work limits without token/page budgets | Complete raw-content form/provenance, human per-content admission and complete relevant-context delivery cannot be inferred | Actual input corpus and admission provenance; observed selection/context accounting |