memory-search

The project's store (topics/, incidents/, investigations/ at the store root) is queried through memidx.py, never by reading the markdown tree directly. Run these from the MemContinuum checkout (or wherever MEMCONTINUUM_HOME/--project are configured for this project).

Commands

Semantic search with filters — the default entry point:

memidx.py search "QUERY" --mode hybrid --project PROJECT [--area AREA] [--type topic] --json

hybrid (the engine's default) is what this skill recommends too — a ruling's own text is retrievable through either channel, so no single-channel mode is needed. Leave --status off: the no-flag default already keeps active records plus anything that never carries a status of its own at all (a record that was never topic-shaped — a sources/ note, this store's own README, a frontmatter-less inbox/ drop) -- that status-less case is exactly what an explicit --status active would drop, so passing it here would be STRICTER than the default, not merely spelling it out. Pass --status any to widen further, to superseded/historical/declined/provisional records, or name one or more of them explicitly. --authority owner-verbatim/owner-ratified narrows to CONSTRAINT-citable rulings only. search also leaves inbox/ drops (freeform reviewer consults, not yet promoted to a record) out of its results by default; pass --include-inbox on the rare occasion you need to search those too — it works at the default status shown above with no second flag, since a freeform drop with no frontmatter at all is exactly the status-less case that default already covers.

A specific topic's full chain, once you have its id or slug:

memidx.py chain <topic-id-or-slug> --project PROJECT --json

What a file is governed by, before editing it (this is also what the PreToolUse hook runs before an Edit/Write — use it by hand when you are about to change files from the shell, or while planning):

memidx.py for-path path/to/file.ext --project PROJECT --json

Drop --json for a human-readable form: chain/for-path print the compressed chain view; search prints one line per hit -- score, path, type, and title.

Reading the output — three tiers (docs/SCHEMA.md §4)

• CONSTRAINT — may refuse, block, or reverse work. Only a link whose ruling.authority is owner-verbatim or owner-ratified and status: active. Code and tests are a separate constraint channel and need no record.
• HOLD — must pause and revalidate against current source. An incident/constraint with reproducible evidence (repro, commit, failing test), any authority. A confirmed current-source safety violation may block on its own; green tests alone don't clear a HOLD unless they discriminate against the failure.
• CONTEXT — informs only. agent-inference never overrides a fix that code and tests accept by itself.

A topic's current link is whatever is newest with status: active — read the whole chain (newest first) when a reversal or amendment is present; the reason a ruling changed (reason_for_change) matters as much as the ruling itself.

The rule

Name what you found in your report, or state that nothing was found. Don't silently absorb a chain into your own reasoning without citing it, and don't silently skip the search because a task "feels" novel — that's exactly when a prior HOLD or declined approach is most likely to be missed. If search/chain/for-path return nothing, say so explicitly ("no existing decision found on X") rather than proceeding as if the question had never been asked.

Before writing a new helper or file

Before adding a new function, type, or file to a codebase this project indexes, search the code index too — a near-duplicate of what you're about to write may already exist:

memidx.py code-search "INTENT PHRASE" --project PROJECT --mode hybrid --json

code-search returns code only by default; --tests adds test code, --all both (they select among chunks already indexed; Swift Tests directories are pruned unless wired as their own code root). Each code hit carries tested_by (up to three test chunks whose name, signature, doc or indexed body excerpt names its symbol) and a role. Phrase the query as the intent ("write a debug PNG", "embed a view in a scroll box", "hash a file's contents"), not as a symbol name — the index is built to match on that. A hit whose concept_id is set means a decision record governs that code; check it (memidx.py why <path>) before working around or duplicating it. The --json output wraps hits in an envelope: state is uninitialized (exit non-zero, [] is a refusal not "nothing found"), stale (a root's content changed since the last code-reindex), degraded (something is incomplete, nothing changed), or metadata-current/current (nothing changed; current only under --verify-content, which hashes every file to prove it — the default metadata-current did not); code_roots, changed/failed/not_indexed, and embedding_mode carry the detail; code-search heals a stale/degraded index itself, once, before answering. Confirm the code index is initialized and not stale before trusting a "nothing found" — real evidence only under current; under metadata-current it is honest uncertainty, not proof of absence — pass --verify-content when that matters, then run code-search; name relevant hits in your report, or say none — the same rule as the markdown search above: silently skipping this check is exactly how a second, slightly different writeDebugImage gets written next to the first one.