ai-memex-cli boundary

Type: agentic-system-analyses/types/agentic-system-boundary.md

Boundary and evidence

This run analyses how ai-memex-cli captures source material, maintains durable Markdown knowledge and makes that knowledge available to later agent sessions. The target is the repository's knowledge system, not only its library functions or deterministic CLI. It is in scope as a memory/knowledge/context-engineering system: the shipped skills serve model-driven workflows, and command implementations launch external agents for semantic work. The package exposes memex through dist/cli.js; its source driver imports the capture, maintenance, retrieval and integration commands (SRC-1, package.json, src/cli.ts).

The boundary is whole-system at the repository level. It includes the CLI driver, its state-producing and state-consuming commands, agent adapters, installable skills and workflow references, vault schema and page templates, generated integration instructions, maintenance loop and source-level tests. The shipped CLI and skill callers are part of the target even though actual model inference runs in an external agent executable. The separation between management and use is explicit in the shipped skill (SRC-1, templates/skills/ai-memex/SKILL.md):

This skill makes the agent the main interface for maintaining an LLM wiki. Use the memex CLI as a deterministic toolbox, not as the source of semantic judgment. --- templates/skills/ai-memex/SKILL.md

Practical rule: if a workflow modifies wiki pages, prefer the CLI command that spawns a sub-agent. If it only reads or summarizes, do it in the current session. --- templates/skills/ai-memex/SKILL.md

The initial trace covers both directions of the retention loop:

• Produce and retain: initialization writes the vault layout and schema; fetch writes captured Markdown under raw/<scene>/; distillation has mechanical session-conversion and delegated-agent paths; ingestion reads the live vault schema and index, then launches an agent with the vault as its working directory. Its prompt directs updates to wiki pages, index.md and log.md while preserving raw sources (SRC-1, src/commands/init.ts, src/commands/fetch.ts, src/commands/distill.ts, src/commands/ingest.ts). These are implementation paths and prompt contracts, not evidence that an agent complied.
• Maintain and check: deterministic lint constructs reports from reloaded wiki pages; optional semantic lint delegates to an agent. The watcher calls the ingest/lint loop, including file-event and periodic health-check paths. The loop feeds prior mechanical findings into subsequent ingestion and can refresh registered project contexts after a clean report (SRC-1, src/commands/lint.ts, src/commands/watch.ts, src/core/ingest-lint-loop.ts). A clean mechanical report does not establish factual correctness.
• Reload and consume: search reloads retained wiki/raw content; inject reads schema includes and optionally keyword-selected wiki pages; glob writes a project-local projection. Context installation and refresh reload the wiki index and render instructions/digests into host files. The query reference directs the current outer agent to search, read pages and answer with citations (SRC-1, src/commands/search.ts, src/commands/inject.ts, src/commands/glob.ts, src/commands/context.ts, src/core/context-block.ts, templates/skills/ai-memex/references/query-workflow.md). Host execution of those instructions is not observed here.

The ingestion prompt makes the intended durable writes explicit (SRC-1, src/commands/ingest.ts):

1. Read all source files in the target.
2. Create or update wiki pages following the schema above.
3. Never modify or delete raw source files.
4. After writing pages, update ${vault}/index.md and ${vault}/log.md as instructed in the schema. --- src/commands/ingest.ts

Repository coverage

The table classifies every top-level tracked file or directory. “Included” means within the functional evidence boundary, not exhaustively inspected. Named paths identify material responsibilities. Initial inspection consists of source reads and bounded tree/call-site searches; available but unread material is not an access gap.

External dependencies and evidence limits

External coding-agent executables and their model providers are dependencies, not enclosed runtimes attributed to this target. Adapter profiles select executable names, prompt delivery and flags (SRC-1, src/core/agent-adapter.ts). Their internal planning, tool permissions, model behavior and provider persistence cannot be concluded from those invocation definitions. Ripgrep and optional qmd are retrieval dependencies; network retrieval relies on external websites/search responses and imported libraries (SRC-1, src/commands/search.ts, src/commands/fetch.ts, package.json). Their live contents and service behavior are not frozen evidence.

No user vault, agent-session transcript, provider response, execution trace or causal experiment was supplied. The Windows vault paths embedded in repository bootstrap files do not grant access to that vault. These gaps prevent conclusions about actual retained knowledge, semantic accuracy, cross-session improvement, runtime compatibility or successful autonomous maintenance. Tests inspected here use stubbed ingest/lint functions; they support understanding the intended control contract, not observed model performance (SRC-1, tests/core/ingest-lint-loop.test.ts). The target's Git-backed description does not itself establish an automatic commit policy.

The evidence is frozen to Git commit 4adb274a33cbb9ff0f03a6aeeb9c8d2525a5ac3f, verified as the checkout HEAD with no reported working-tree changes. The applicability cutoff is 2026-10-15, the supplied run date; this does not imply that the revision was current upstream on that date. Only this repository at this commit supplies evidence. No linked upstream pattern, dependency repository, live web page or later release is included. Overall evidence tier: code-grounded, with implementation and separately attributed doctrine/design; no observed-run or causal-experiment layer.

Source register