MemWiki synthesis
Type: agentic-system-analyses/types/agentic-system-synthesis.md
Bounded synthesis
MemWiki is a project-memory protocol installed by a file-writing CLI, not a model runtime or an executable memory service. This account covers the material shipped setup, maintenance and consumption paths in https://github.com/hereisSwapnil/memwiki at commit 8034a3da991ac2639b87875172a9572903ecf1d5, with an analysis cutoff of 2026-10-14. SRC-1 supplies inspected implementation, doctrine/design and an attributed ingestion report. No target execution, host-model trace or causal comparison was supplied. The whole-system boundary includes the host-facing instructions and retained wiki, but excludes external host implementations and published npm bytes. See boundary and evidence and the Source register.
Install the protocol, then enter through the host
The operator runs initialization in a project directory. The inspected TypeScript CLI accepts init and update, creates directories, and iterates fixed templates for three host hooks, AGENTS.md and seven wiki files. Initialization writes absent files and skips existing ones, even if an existing hook does not route to MemWiki. The invoking process's working directory identifies the project; its filesystem permissions bound the effects. Console output terminates the pass. There is no model invocation, transaction or rollback; rerunning can fill missing files but does not repair existing content. These are wired implementation conclusions, not observed runs. SRC-1, src/index.ts; RT-RTE-cli-init.
An external coding host must discover a hook, expose file tools and follow the installed instructions. The foreground agent retains decision ownership. MemWiki assigns no separate reviewer or worker-return channel. Its protocol is interpreted natural language, not a permission check. The Claude hook requests primary-source priority for the protocol and hot cache; actual priority and compliance depend on the host. SRC-1, CLAUDE.md, AGENTS.md; RT-OBJ-protocol-files, RT-BAP-host-protocol, RT-CMP-host-agent.
Reload selected context and turn work into durable files
At session entry the agent is instructed to read .memory/wiki/hot.md first. If that is insufficient, it consults the index and relevant domain pages. Hot supplies current state and next steps; the index supplies navigation; stack, patterns, bugs and decisions supply project reference material, conventions and some architectural rationale. This is requested pull reading by the foreground consumer, not independent relevant-part injection or an implemented retrieval ranker. The roughly 500-word hot-cache description is guidance, not a hard budget. Conditional domain reading also means that retained rationale need not reach every later decision. SRC-1, AGENTS.md, README.md; RT-RTE-agent-session, MEM-OBJ-hot, MEM-OBJ-index, MEM-OBJ-domain-pages, MEM-OBJ-decisions.
During work, the protocol asks the same agent to record patterns, complex bug fixes and significant decisions, and to synthesize human-supplied raw documents into wiki pages. New pages must be indexed. Before session end, major-task completion or impending context loss, the agent must refresh hot context and append a timestamped work summary. User launch is not per-sentence approval: the model selects and admits the derived content through host tools. Raw originals have requested immutability, not filesystem-enforced protection. Their actual payloads and provenance are unresolved because none is tracked in this source. SRC-1, AGENTS.md, src/index.ts; RT-RTE-agent-session, MEM-OBJ-raw, MEM-OBJ-log.
The representative workflow is chat-requested /memwiki-ingest: read the installed context, scan repository material, populate stack/patterns/decisions, then perform the same session-end hot/log obligations. Delivery is a durable file handoff to a later coding session, not a direct agent message or machine-checkable acknowledgment. The log reports ingestion and populated pages exist, but neither establishes a complete scan, accurate synthesis or subsequent behavioral dependence. Supplying the chat command to the CLI instead is rejected; operator recovery requires using the host chat channel. Interrupted agent work has no assigned checkpoint or rollback procedure in this protocol. SRC-1, AGENTS.md, src/index.ts, .memory/wiki/log.md; MEM-RTE-repository-ingest, workflow account.
Maintain context without confusing maintenance with warrant
/memwiki-lint is a separate optional health request, not an ingestion gate. It asks the agent to diagnose outdated information, missing hot context and empty pages, then propose fixes. The README's promise of fixing is stronger than that protocol; admission of proposed edits remains unresolved. /memwiki-fold requests condensation of older log entries while preserving crucial history. This permits textual replacement despite append-only language. Neither a preservation audit nor removal of already loaded host turns is specified. SRC-1, AGENTS.md, README.md; MEM-RTE-wiki-lint, MEM-RTE-log-fold.
Retention and requested reliance do not establish accepted knowledge. Within the inspected CLI, hooks, protocol and README, no explicit evidence-consuming acceptance decision or mandatory withdrawal gate was found. Lint is a checking request, but its advisory proposal does not supply that missing disposition. Concrete retained inconsistencies qualify the primary-truth language: domain prose claims no runtime dependencies while package metadata declares memwiki; the log reports landscape/competitor links that the frozen index lacks. These are bounded artifact discrepancies, not evidence that lint ran and failed or that all retained claims are false. SRC-1, package.json, .memory/wiki/stack.md, .memory/wiki/log.md, .memory/wiki/index.md; EPI-ABS-acceptance-gate, EPI-CLM-primary-truth, MEM-OBJ-domain-pages, MEM-OBJ-log.
Protocol update is a different change mechanism: fixed templates overwrite four hook/protocol files without merging local customizations. Existing wiki files are skipped under ordinary single-process filesystem assumptions, but missing wiki pages are created. This preserves existing content, including stale assertions; it neither validates nor refreshes that content. Direct execution of tracked src/index.js is an alternate init-only route with older instructions, not the declared package bin. SRC-1, src/index.ts, src/index.js, package.json; RT-RTE-protocol-update, RT-CLM-safe-update, RT-RTE-legacy-init.
What the retained loop supports
The strongest learning-related contribution is a declared work-to-durable-context loop: agent-selected hot/log summaries, and further folding of retained history, have identified later-consumer roles. The accepted profile verification supports trace learning at claimed route level, independently of whether capacity improved. It also identifies additional domain-discovery paths missing from the profile's per-route inventory. Original trace types remain indeterminate; timestamps do not establish session logs or tool traces. SRC-1, AGENTS.md, .memory/wiki/log.md; RT-RTE-agent-session, MEM-RTE-log-fold, profile verification.
Whether criticism of consumed guidance improved future action remains unestablished: there is no supplied criticism-to-revision-to-later-action comparison. The runtime does not establish theory-builder conditions 1–4. Retained rationale is inspectable, but it does not by itself establish a formulated-theory assessment, content-dependent consumption, content-directed criticism or iteration. Reflection, autonomous maintenance and self-improvement likewise remain independently unestablished, not evidenced absent. For exercised self-improvement, the missing link is especially later operation causally depending on the evidence-responsive change; file availability or instruction replacement does not close it. SRC-1, AGENTS.md, src/index.ts; MEM-OBJ-decisions, RT-RTE-agent-session, epistemic bounded conclusion.
For a cooperating coding agent that can discover instructions and read/edit local files, this architecture affords continuity through selected context and retained reasons. For a scenario requiring guaranteed persistence, validated project facts or automatic recovery, the inspected mechanisms support no such guarantee. Host traces showing reads, writes and subsequent decisions would establish exercised paths; paired source/wiki and pre/post-fold artifacts would constrain transformation fidelity. An explicit acceptance/withdrawal consumer would change the warrant assessment. A controlled comparison varying retained context would be needed to attribute benefit to this memory mechanism rather than to other host behavior. No source files were read during synthesis.
Limitations
| Limitation | Affected SRC IDs and record citations | Inspected boundary | Conclusion prevented | Evidence that would resolve it |
|---|---|---|---|---|
| Unresolved index enumeration conflict: workflow says seven listed pages; memory describes six outgoing links, and the epistemic quotation contains six. Both positions remain recorded. | SRC-1; MEM-OBJ-index, EPI-OBJ-index, workflow delivery, report-verification limit | Frozen .memory/wiki/index.md and supplied delivery accounts |
A consistent exact outgoing-link count across the reports; conditional routing and the missing landscape/competitor links remain supported | Correction of the workflow count against the retained index quotation |
| Profile omits separate trace-learning assessment of work-discovery writes into patterns, bugs and decisions. | SRC-1; RT-RTE-agent-session, MEM-OBJ-domain-pages, MEM-OBJ-decisions, profile-verification limit | Accepted work-write and domain-reader records | Complete per-route trace-learning and trace-origin comparison; claimed positive existence remains valid | Separate classification of each domain-discovery write and its original activity inputs |
| External host discovery, precedence, approvals, model state and actual consumption are unresolved; no observed run or causal comparison is supplied. | SRC-1; RT-CMP-host-agent, RT-BAP-host-protocol, RT-RTE-agent-session | Shipped hooks/protocol, excluding host implementation | Deployment compliance, activated read-back, demonstrated learning or exercised self-improvement and attributable benefit | Pinned host contracts and inspectable cross-session traces; appropriately controlled outcome comparisons |
| Published executable and dependency bytes are outside the source boundary. | SRC-1; RT-RTE-cli-init, RT-RTE-legacy-init, boundary evidence | Source executables and package/build metadata; no tracked dist or frozen npm artifact | Installed-byte equivalence, successful build and dependency implementation behavior | Frozen distribution/dependency artifacts and build provenance |
| Concrete raw payloads, paired synthesis inputs/outputs and original work traces are missing. | SRC-1; MEM-OBJ-raw, MEM-OBJ-log, MEM-RTE-repository-ingest | Raw-source instructions and retained wiki candidates | Complete input form/provenance, sentence lineage, full ingestion coverage, semantic relation and controlled trace origin | Original inputs linked to transformation and consumer traces |
| Lint proposals are not an established retained correction; fold fidelity is untested. | SRC-1; MEM-RTE-wiki-lint, MEM-RTE-log-fold, EPI-ABS-acceptance-gate | Maintenance doctrine and executable dispatch | Guaranteed repair, withdrawal of stale assertions, semantic preservation or accepted knowledge | Recorded diagnosis, explicit disposition and later consumer; paired fold artifacts with preservation checks |
| Retained reports conflict with surviving artifacts, and hot freshness is unresolved. | SRC-1; MEM-OBJ-log, MEM-OBJ-index, MEM-OBJ-domain-pages, MEM-OBJ-hot | Frozen wiki and package metadata | Authenticated production history, reliable dependency-free claims and successful hot freshness | Source-linked history explaining missing pages/links, reconciled dependency claims and a dated task-state check |