The architectural bet is that no model belongs in the write path: ingestion and recall are deterministic, agents propose and you approve, and every memory is a Markdown file you can grep or delete by hand. That also means the useful failure mode is legible — a weak or semantic-only match comes back labelled as such, so the agent asks instead of asserting.
A memory layer that several agents share. Raw notes, transcripts and docs become an inspectable Markdown wiki; explicit "remember this" requests become reviewable memories; agents retrieve compact, source-backed packets through the CLI, MCP, skills or a local viewer rather than dumping the wiki into a chat window.
- Readiness, schema state, validation and safe next actions before anything is written — `status`
- The single read path — startup briefs, answer-ready packets, wiki search, graph context and follow-up actions, each memory carrying a confidence label of `strong`, `moderate` or `weak` and a match type of `lexical`, `semantic` or `hybrid` — `recall`
- Durable memory written only after explicit approval, with duplicate and conflict checks, provenance, visibility, and optional `review_after` and `expires_at` — `remember`
- Raw sources turned into wiki pages, with stale-ingest detection when a file changed after it was indexed — `ingest`
- The memory inbox, audit, archive, restore and forget workflows, plus a read-only consolidation plan applied per action with approval — `review`
- Backup, migration, graph export, rebuilds and advanced maintenance — `admin`
- Point-in-time answers: asking what was true in March reconstructs from dated files and their supersede lineage, matched by a regex and a calendar rather than a model
No account, no key and no hosted backend — everything is local files, and the installed CLI, MCP server and viewer carry no telemetry. Install with Homebrew, then `lnk setup`, which detects Claude Code, Codex, Cursor, Windsurf, Zed, Kiro and Gemini CLI and wires them in one pass. For an MCP-only install, `pip install link-mcp` and launch `python3 -m link_mcp --wiki ~/link/wiki --surface slim`; the server.json publishes `link-mcp` at 2.3.0 over stdio. `--surface slim` is the recommended six-tool surface, `--surface full` keeps the compatibility set. Paraphrase matching needs the optional `link-mcp[semantic]` extra and a one-time model fetch you approve; recall itself never touches the network. On Homebrew Python, `externally-managed-environment` means install into a dedicated venv.
Build from source — clone the repository and build it, then point your client at the binary