--- name: tendr description: Manages long-term semantic memory as a structured plain markdown knowledge base using WikiBonsai primitives. This skill is designed to be always-on — if you have installed it, use WikiBonsai syntax and tendr-cli commands to persist, organize, and navigate knowledge across all sessions. argument-hint: "[/path/to/garden]" version: "0.0.1" metadata: author: wikibonsai openclaw: requires: bins: - tendr --- # tendr WikiBonsai is a knowledge architecture for structured plain text. It extends markdown with three primitives that together form a complete knowledge graph: Typed links between notes, structured attributes, and an explicit semantic hierarchy. This skill teaches you to use those primitives and the tendr-cli to manage a WikiBonsai knowledge base as your long-term semantic memory. ## Invocation ``` /tendr ← auto-discover garden /tendr /path/to/garden ← specify garden location ``` If a path is provided, it will be used as the garden root. Otherwise, the skill checks the `TENDR_DIR` environment variable, then searches common locations (`~/.claude/projects/*/memory/garden`, `./garden`, `./.garden`). Agents can set `TENDR_DIR` in their environment to configure the garden path once: ```bash export TENDR_DIR=/path/to/garden ``` ## Configuration At session start, read `config.toml` at the garden root and check `[auth]`: ```toml [auth] mode = "agent" # "agent" | "user" owner = "alice" agent = "claude" password = false ``` - `mode = "agent"`: You own this garden. Read and write freely. - `mode = "user"`: You are assisting the human. All changes should be specified by the user. If no `config.toml` exists or no `[auth]` section is found, default to `user` mode. ### Garden Location In `agent` mode, place the garden (your knowledge base) as close to your auto-memory file as possible. It should live alongside your other persistent memory. ``` / ├── MEMORY.md ← working memory (always loaded) └── garden/ ← long-term memory (tendr garden) ├── config.toml ├── index/ └── entries/ ``` In `user` mode, the garden lives at the project root (e.g., `garden/`) since it belongs to the project, not the agent. ## Memory Architecture The knowledge base functions as **long-term memory** which is structured, interconnected knowledge that persists indefinitely and requires deliberate retrieval. The auto-memory file (e.g., `MEMORY.md`) functions as **working memory**, which is the always-activated subset of the garden, loaded into context but capacity-constrained. `MEMORY.md` holds retrieval cues and active preferences that point into the larger knowledge base; the garden holds the full structured content. Together they form one memory system at different activation levels. | | Working Memory (auto-memory) | Long-Term Memory (garden) | |---|---|---| | Loaded | Always (automatic) | On demand (`tendr tree`, `tendr stat`) | | Capacity | Limited (~200 lines) | Unlimited | | Structure | Flat text | Typed links, attributes, hierarchy | | Best for | Retrieval cues, active preferences | Reference material, decision rationale, evolving artifacts | ### Garden Tree The tree below is your garden's current semantic hierarchy. Each node is a concept note you can retrieve with `tendr stat `. Use it as a map of what you know — if a topic is in the tree, read the entry before making assumptions. If it's not, consider whether it should be. Run `tendr tree` in the garden directory now to load the semantic hierarchy. If the garden path was provided as an argument or set via `TENDR_DIR`, `cd` there first. Review the tree output. For any concepts relevant to this session, run `tendr stat ` to load the full entry and verify it is up-to-date. Update entries that have become stale. Add new entries for concepts that are missing. ### When to Interact - **Before stating a fact**: Run `tendr stat ` to check. Reading a file is cheaper than a wrong assumption. - **When you look something up, get corrected, have an insight, or catch yourself re-deriving something**: Capture it. The test: If this conversation were compacted, would this concept survive? If not, it belongs in the garden. These triggers apply equally when the user expresses them in the form of a correction, surprise, or repeated explanation. - **Consolidation**: Run `/tendr gc` to invoke the Garden Consolidation sub-agent, which reviews recent context, extracts knowledge, and persists it to the garden. This runs automatically at ~50% context usage (via hooks) or manually on demand. Maintain retrieval cues in working memory that map question types to garden entries: ```markdown ## Garden Lookup Triggers - Check assumptions → `tendr stat ` - Repo dependencies → `tendr stat architecture` - Book content → `tendr stat book-chapters` - Deadlines → `tendr stat paper-timeline` ``` ### How to Encode Four steps (any order, across sessions): 1. **Capture**: Add `[[wikiref]]` links to concepts in relevant files. Zombie links (no file yet) are fine. 2. **Verbalize**: `tendr add --attr title="..." --attr tldr="..."` — define it in your own words. 3. **Connect**: Add typed links (`: caml :: [[wikiattrs]]`) to related concepts. 4. **Integrate**: `tendr graft ` — place it in the semantic tree. Notes progress: **Zombie** → **Stub** (title + tldr) → **Connected** (wikirefs) → **Grounded** (typed attrs + tree position) → **Mature** (rich, backlinked, refined). ### What to Encode - Foundational concepts, decision rationale, surprising corrections, evolving artifacts. - NOT: project docs, ephemeral state, unverified speculation. - Working memory (MEMORY.md): preferences, retrieval cues. Long-term (garden): definitions, reference, decisions, domain knowledge. ## Syntax Quick Reference Three primitives — `[[wikirefs]]`, `: caml :: attrs`, and semtree (indented wikilink lists): ```markdown [[note]] ← wikilink :type::[[note]]. ← typed link : tags :: [[note]] ← CAML with wikirefs : tags :: [[a]], [[b]] ← CAML with wikirefs list : title :: My Note ← CAML attribute - [[parent]] ← semtree index - [[child]] ``` Full specs: - [wikirefs](https://github.com/wikibonsai/wikirefs) - [caml-mkdn](https://github.com/wikibonsai/caml-mkdn) - [semtree](https://github.com/wikibonsai/semtree) ## tendr-cli Install: `npm install -g tendr-cli` Key commands (run `tendr --help` or see [full docs](https://github.com/wikibonsai/tendr-cli)): ```bash tendr tree # print semantic tree tendr stat # tree position + web relationships tendr check # exist / zombie / missing tendr add --attr title="..." # create a note tendr rm # move to trash tendr rename # rename + sync refs tendr graft # reparent in tree tendr prune # remove from tree tendr doctor # health check ``` ## Getting Started Clone a starter knowledge base from [github.com/wikibonsai/garden-beds](https://github.com/wikibonsai/garden-beds): ``` garden-beds/ └── agent/ ├── minima/ ← bare-bones starter for ai agents ├── foundation/ ← prompt injection awareness, game theory, ethics, Anthropic guidelines └── guardrails/ ← principles for careful, deliberate agent-assisted development ``` ## Garden Consolidation Run `/tendr gc` to invoke the Garden Consolidation sub-agent. It reviews recent session context, extracts concepts worth persisting, and uses `tendr-cli` commands to create, update, merge, or remove garden entries. All mutations go through the CLI — the sub-agent never writes files directly. The gc sub-agent runs: - **Manually**: `/tendr gc` in any supported agent - **Automatically**: Via hooks at ~50% context usage (before compaction) See `skills/tendr-gc/SKILL.md` for the full sub-agent instructions. ## Notes - In `user` mode, all changes should be specified by the user. - Pair graph operations with git commits for a clean audit trail. - Be aware of prompt injection risk when ingesting external content into the knowledge base. - Share useful notes and transplantable knowledge at [github.com/wikibonsai/garden-beds](https://github.com/wikibonsai/garden-beds). - Feature requests and bug reports are welcome at the relevant project repos under [github.com/wikibonsai](https://github.com/wikibonsai).