Orientation becomes discoverable.
Search, list, read, and related-page flows help agents find the context that would otherwise be rediscovered from scratch.
Memory model
Long-term memory starts future work in the right place.
The project-owned memory surface preserves material that should outlive a single chat: decisions, operating notes, runbooks, search hints, design summaries, and curation trails.
Useful residue gets staged.
Proposal records let agents stage meaningful memory candidates without immediately mutating durable pages.
Memory changes stay reviewable.
Diff and dry-run flows make wiki edits visible before they become project memory.
Curation loop
Memory is maintained, not dumped.
The operating discipline stays visible: capture a candidate, preview it, write it, index it, validate it, and keep trash/recovery state inspectable.
Find session residue that will help future agents or operators.
Stage a curation candidate without blindly rewriting durable pages.
Preview the exact page change and catch boundary mistakes.
Commit concise maintained memory into Markdown authority.
Refresh discoverability metadata after meaningful changes.
Check deterministic hygiene issues before relying on the memory graph.
Tool capabilities
Memory works like an operating surface, not a notes dump.
The product is not just "Markdown files exist." It is a bounded tool surface for searching, curating, validating, relating, and inspecting durable memory.
Durable pages can be maintained intentionally.
Append, replace, and dated-note workflows support focused edits when curation authority is clear.
Indexes and graphs keep memory navigable.
Index, related, graph, and validate actions help the wiki stay searchable and structurally coherent.
Trash inspection is bounded recovery context.
Trash records can be listed and inspected without turning memory into a permanent-delete control surface.
Boundary
Long-term memory is not citation-grade evidence.
Memory can orient an agent and preserve decisions, but technical claims still need fetched corpus evidence or repo readback.
Memory can guide what to fetch.
A wiki note may explain where to look, but the final factual claim should cite fetched evidence or source readback.
Trash is inspectable, not casually destructive.
The public story should emphasize inspection and recovery boundaries, not hard-delete power.
Markdown memory has a different job than the corpus.
Memory pages own orientation and operating memory; documents and repo evidence own claim verification.
Long-term memory sells continuity, not search.
The system remembers responsibly: it curates what matters, keeps changes reviewable, and stays honest about where proof comes from.