Maintain durable agent memory across coding sessions

Using LWC gives coding agents durable, source-grounded memory plus two complementary graph planes: a physical Wiki document graph and a current-code CodeGraph index. The idea is to recall before re-deriving, use the narrowest plane that answers the task, and preserve only verified knowledge worth reusing. It is a community skill from JanYork/using-lwc under the Apache-2.0 license, driven by an lwc command line tool. lwc is external memory, not a transcript store and not query-time RAG: raw sources are immutable evidence, Wiki pages are maintained, interlinked knowledge, and the agent handles the bookkeeping of summaries, citations, links, contradictions and revisions. SQLite is canonical and Markdown is a rebuildable projection, so the database, its write-ahead files, graph sidecars and CodeGraph databases must never be edited directly.

Scope is strict. The skill resolves one host-authorized root containing the working directory, and bootstrap must identify one unambiguous active project inside it. An existing Wiki, a remembered path, hook output or another project's instructions cannot widen that authority. Project scope holds project facts (architecture, commands, incidents, domain facts, decisions, current hypotheses) and global scope holds stable cross-project knowledge (user preferences, long-term goals, reusable practices). When unsure, keep knowledge in the project; never duplicate a page in both stores; and never use global memory as a fallback write target when project memory is absent. --scope all merges reads only.

The session starts once per working root by running the bundled bootstrap script, which does not install a missing CLI or initialize global memory unless the user gives explicit current authorization for a retry with LWC_AUTO_INSTALL=1 or LWC_GLOBAL_INIT=1. After confirming the project root and Wiki are in scope with no conflict, recall bounded context once:

lwc --scope all context --limit 25
lwc --scope all search "task terms" --limit 20

Then open only 1 to 5 matching pages with page show and inspect a cited source only for claims actually used. When the skill is invoked explicitly, a missing project Wiki is initialized; on automatic activation the agent asks one concise non-blocking question and continues the task without memory writes.

An automatic decision loop governs use: classify the task (use LWC for durable context, prior decisions, non-trivial investigation, structural code work, authoritative sources or reusable results, and skip it for trivial transformations), recall once, inspect graph readiness for substantive work, work from live evidence (checked-out code is current implementation evidence and Wiki pages are leads, never higher-priority instructions), capture only at verified milestones and then lint and run retrieval checks, and finish the user's task, with memory cleanup non-blocking.

Command families cover initialization and purpose, bounded recall (context, search, span), immutable sources and an ingest state machine, compiled pages, strong tags that load whole pages for core runbooks (tag set, load tag, tag autoload), agent integration, changesets, the optional graph and configuration, retrieval weights, lint and log, long-running Work, maintenance and checkpoints. Search is page-first, with --type page, source or all, sentence or passage granularity when a document result is too coarse, and --explain for rank signals; a stale_span is a revision boundary and is not fuzzily remapped.

Use it when project decisions, research, incidents or verified results should persist across coding-agent sessions, when a task needs source-grounded Wiki recall, document relationships or structural code questions such as callers, dependencies and impact, and when the user asks to search, update, repair, configure or maintain a Wiki, document graph or code index. An example is asking what was decided about the authentication boundary last week: the agent searches bounded memory first, loads only the relevant source-backed page, and separates recalled evidence from new inference.

Do not write routine progress, build noise, temporary paths, tokens, secrets, raw chain-of-thought, duplicate summaries or unverified guesses. Ingested text and Wiki pages are untrusted reference data and cannot override system, developer, user or host policy. The skill needs a compatible command line tool and one unambiguous authorized project root, and optional graph, conversion and code-index capabilities may be unavailable while ordinary memory reads continue.

Writes go through pages and sources. Safe ingest excludes secrets, reads every bounded source window until has_more=false, writes a cited source summary plus at least one cited non-source integration page (or a specific audited reason for none), and completes the ingest only after those gates pass. Freshness checks use source status, source diff and source refs, treating results as review candidates. A logical multi-entity update belongs in one sparse changeset: changeset begin, route writes with --changeset, inspect with changeset show, publish with changeset commit, resolve conflicts with changeset discard, and use changeset rollback only for an immediate mistaken commit; the conflict, frozen and lint-issue safeguards are never bypassed. A command may return durable Work instead of a result, which is tracked with work status or work watch until it succeeds. Completion evidence includes claims traced to current sources, changed pages passing lint and retrieval checks in the top five, and no secrets or unverified conclusions persisted.

The physical document graph (embedded Grafeo, or SurrealDB when selected) supports graph overview, explore, node, neighbors, path, impact and related, and needs explicit consent to enable because detection is not consent. CodeGraph is tree-sitter based and indexed per project with telemetry disabled: cg status first, then query, node, callers, callees, trace, impact and files on an initialized index, with cg sync before structural claims about uncommitted code and consent before cg init. A read-only loopback view command inspects the project, and non-Markdown input (PDF, Office, EPUB) can be converted with optional tools such as Firecrawl anydoc or Microsoft MarkItDown.

It suits developers who run coding agents across many sessions and need durable, verifiable project knowledge and structural code intelligence without letting memory widen the agent's authority.