Tessellum

Typed atomic notes in a graph — a Zettelkasten that scales.

Knowledge construction for humans and agents, built on six architectural pillars: Zettelkasten, PARA, Building Blocks, Epistemic Functions, Dialectic Knowledge System (DKS), and CQRS.

Tessellum is a knowledge-construction system, not an agent-memory store. The unit of work is a typed atomic note — a tessellum, a small mosaic tile — that carries one epistemic claim. You write tessellae; Tessellum indexes them, retrieves them with hybrid BM25 + vector search, lets you grow Folgezettel trails that record how thinking developed, and runs a closed-loop Dialectic Knowledge System that updates warrants from observed disagreement. The architecture is CQRS: a typed prescriptive substrate (what you author) and a computational descriptive retrieval layer (what queries return) — read-side and write-side cleanly separated.

Status

v1.0.0 — every engine subsystem shipped, headlined by the Composer v4 dynamic-workflow runtime. Suite: 1152 tests.

• Composer — a typed-contract pipeline runtime. Skill canonical + .pipeline.yaml sidecar → zero-LLM compile → typed DAG → execute either serially (run_pipeline, the byte-identical reference path) or through the v4 self-claiming, wave-parallel dynamic scheduler (run_pipeline_dynamic, opt-in via --dynamic) with a resume manifest, a plan/session/wave gate engine, an error-class + full-jitter retry ladder, run-level budgets, a key-rotating credential pool, a fail-soft context assembler, and a pluggable sign-off approver. Four LLM backends: Mock / Anthropic / Bedrock / Pooled.
• DKS — a shipped closed-loop Dialectic Knowledge System runtime engine (tessellum dks): a 7-component cycle (observation → N arguments → contradicts edges → counter → pattern → revised warrant) over the Building-Block graph, multi-perspective Dung argumentation, confidence gating, warrant persistence, and a second-order meta-DKS that mutates the BB schema itself.
• Retrieval — BM25 (FTS5) + dense (sqlite-vec) + hybrid RRF + best-first BFS + metadata filter, over the indexed vault.
• Indexer — vault → one SQLite DB (notes + note_links + FTS5 + sqlite-vec, all-MiniLM-L6-v2 384-d).
• Format — closed-enum YAML validator + parser + BB-graph-aware link checker. BB — the 8-type, versioned, event-sourced ontology (source of truth).
• Interfaces — an 11-command CLI (init / format / capture / index / search / filter / fz / bb / composer / dks / mcp) and a shipped MCP stdio server (tessellum mcp serve, 7 tools).

See CHANGELOG for the per-release ship list, and docs/ for the architecture + per-module design reference.

The Six Pillars

Two supporting concepts that bridge the pillars (also shipped as term notes):

What Tessellum Is Not

Quick Start

pip install tessellum

# 1. Scaffold a new vault (templates + seed term + master TOC)
tessellum init ~/my-vault
cd ~/my-vault

# 2. Capture your first typed atomic note — 14 flavors available
tessellum capture concept page_rank        # creates resources/term_dictionary/term_page_rank.md
tessellum capture skill my_skill           # creates skill_*.md + paired skill_*.pipeline.yaml
tessellum capture code_snippet my_algo     # creates resources/code_snippets/snippet_*.md
tessellum capture code_repo my_repo        # creates areas/code_repos/repo_*.md
tessellum capture --help                   # full flavor list

# 3. Validate format (closed-enum YAML spec)
tessellum format check .

# 4. Index the vault (notes + links + FTS5 + sentence-transformer embeddings)
tessellum index build

# 5. Retrieve — hybrid RRF default; --bm25 / --dense / --bfs for explicit strategy
tessellum search "graph traversal"
tessellum search --bm25 "PageRank"          # lexical only
tessellum search --bfs term_page_rank.md    # graph traversal from a seed
tessellum filter --tag concept --bb model   # direct metadata filter (tags / BB / status / dates)

# 6. Compose — typed-contract runtime for skill-driven workflows
tessellum composer validate vault/resources/skills/                          # all skills
tessellum composer compile  vault/resources/skills/skill_my_skill.md         # to typed DAG (zero LLM)
tessellum composer run      skill_my_skill.md                                # serial, mock backend (default)
tessellum composer run      skill_my_skill.md --backend anthropic            # real Claude (pip install tessellum[agent])
tessellum composer run      skill_my_skill.md --backend bedrock              # Anthropic-on-Bedrock (AWS creds)
tessellum composer run      skill_my_skill.md --dynamic --workers 8 \        # v4 self-claiming parallel scheduler
    --manifest run.json --close-gate --wave-gate --max-invocations 200       #   + resume manifest, gates, budget
tessellum composer batch    jobs.json --parallelism 8                        # parallel multi-skill
tessellum composer eval     scenarios/  --judge-backend anthropic            # structural assertions + LLMJudge rubric

# 7. DKS — run the Dialectic Knowledge System engine over observations
tessellum dks observations.jsonl --perspectives a,b,c                        # multi-cycle dialectic (N>2 → Dung)
tessellum dks --report                                                       # inter-cycle telemetry
tessellum dks --meta --apply                                                 # second-order: mutate the BB schema

# 8. Serve tools to an agent over MCP (pip install tessellum[mcp])
tessellum mcp serve                                                          # stdio server, 7 tools

tessellum --version prints the version + capability banner.

Architecture

                    ┌──────────────────────────────────────┐
                    │  vault/  (markdown + YAML)           │
                    │  System P — typed substrate          │
                    │   • 8 BB types × ~80 sub-kinds       │
                    │   • PARA categories                  │
                    │   • Folgezettel trails               │
                    └──────────────────┬───────────────────┘
                                       │ indexed
                                       ▼
                    ┌──────────────────────────────────────┐
                    │  data/databases/   (one .db file)    │
                    │  SQLite + sqlite-vec + FTS5          │
                    └──────────────────┬───────────────────┘
                                       │ queried
                                       ▼
                    ┌──────────────────────────────────────┐
                    │  src/tessellum/retrieval/            │
                    │  System D — descriptive retrieval    │
                    │   • Hybrid BM25 + vector via RRF     │
                    │   • Best-first BFS + metadata filter │
                    └──────────────────┬───────────────────┘
                                       │ exposed
                                       ▼
                    ┌──────────────────────────────────────┐
                    │  Interfaces + runtimes               │
                    │   • CLI (11 cmds): init/format/      │
                    │     capture/index/search/filter/fz/  │
                    │     bb/composer/dks/mcp              │
                    │   • Composer v4: skill canonicals →  │
                    │     typed DAGs, serial or --dynamic  │
                    │     self-claiming scheduler;         │
                    │     backends Mock/Anthropic/Bedrock  │
                    │   • DKS engine: closed-loop dialectic│
                    │     over the BB graph (+ meta-DKS)   │
                    │   • MCP stdio server (shipped, 7     │
                    │     tools) + a composer-ts/ bridge   │
                    └──────────────────────────────────────┘

See vault/resources/analysis_thoughts/thought_six_pillars_architecture.md for the full pillar-by-pillar deep dive.

The Building Block Ontology

Every tessellum has a building_block field in YAML frontmatter — one of 8 typed roles, each with a defining epistemic function. The typed edges between roles form a versioned, event-sourced schema graph (~16 edges: 8 epistemic + 7 navigation-index + 1 DKS extension) — the source of truth is src/tessellum/bb/, and the second-order meta-DKS can mutate it. This ontology drives the Dialectic Knowledge System. See vault/resources/term_dictionary/term_building_block.md and docs/bb.md.

Folgezettel Trails

Wikilinks tell you what's related. Folgezettel trails tell you how thinking developed — argument → counter → response → reframe → synthesis, encoded in trail IDs (7 → 7a → 7a1 → 7a1a). See vault/resources/term_dictionary/term_folgezettel.md.

Project Structure

The top-level layout maps each folder to a defined CQRS role — System P (capture), System D (retrieval), or governance/runtime that sits outside both. See plans/plan_cqrs_repo_layout.md for the full workflow → folder mapping.

Tessellum/
├── src/tessellum/         Python code — engines for System P (capture) and System D (retrieval)
│   ├── bb/                Building Block ontology — 8-type, versioned, event-sourced schema graph (source of truth)
│   ├── format/            Validator + parser + BB-graph-aware link checker (closed-enum YAML spec)
│   ├── indexer/           Vault → one SQLite DB (notes + note_links + FTS5 + sqlite-vec; all-MiniLM-L6-v2 384-d)
│   ├── retrieval/         BM25 + dense + hybrid RRF + best-first BFS + metadata filter (+ heuristic router)
│   ├── composer/          Composer v4 runtime — compiler + scheduler (serial run_pipeline + dynamic
│   │                      run_pipeline_dynamic) + executor (retry ladder) + manifest + gates + fix +
│   │                      credential_pool + context_assembler + planning + signoff + llm
│   │                      (Mock/Anthropic/Bedrock/Pooled) + batch + eval
│   ├── dks/               Dialectic Knowledge System engine — core + fsm + dung + confidence + persistence + meta/
│   ├── capture.py         14-flavor capture registry (concept, procedure, skill, model, argument,
│   │                      counter_argument, hypothesis, empirical_observation, experiment,
│   │                      navigation, entry_point, acronym_glossary, code_snippet, code_repo)
│   ├── init.py            tessellum init scaffold
│   ├── cli/               Per-subcommand dispatchers (11 commands) wired into argparse
│   ├── mcp/               Shipped MCP stdio server (7 tools) — `tessellum mcp serve`
│   └── data/              Force-included template directory + seed-vault content
├── composer-ts/           TypeScript orchestration bridge (bridge-not-port; shells the Python CLI)
├── docs/                  Architecture + per-module design reference
├── vault/                 Shared substrate — typed atomic notes (Tessellum dogfoods itself)
│   ├── 0_entry_points/    Master TOC + 5 acronym glossaries (statistics, critical thinking,
│   │                      cognitive science, network science, LLMs) + master glossary index
│   ├── resources/
│   │   ├── term_dictionary/   Conceptual primer (BB, FZ, DKS, CQRS, Z, PARA, …)
│   │   ├── how_to/            How-to guides
│   │   ├── analysis_thoughts/ Architecture arguments + FZ trails
│   │   ├── templates/         15 copy-and-fill skeletons (executable spec exemplars)
│   │   ├── skills/            Skill canonical bodies + pipeline sidecars
│   │   ├── code_snippets/     `## Patterns`-format snippet notes (one component or algorithm)
│   │   ├── code_repos/        Repo notes (main + sub-note structure)
│   │   ├── teams/   tools/   faqs/   digest/   papers/
│   └── areas/             Code-repo notes (main + module sub-notes)
├── inbox/                 System P input queue — drop zone for raw incoming (papers, drafts)
├── plans/                 Governance — project-management plans (committed, top-level)
├── data/                  System D build output (gitignored, regenerable: DBs + embeddings)
├── runs/                  Both-system runtime traces (gitignored)
│   ├── capture/           Capture-pipeline traces (reserved; not yet wired)
│   ├── retrieval/         Retrieval evaluation + benchmark traces (reserved; not yet wired)
│   └── composer/          Composer chain run traces (wired by `tessellum composer run/batch`)
├── experiments/           Experiment outputs
├── scripts/               Operational utilities (one-off migrations; not in wheel)
└── tests/                 Test suite (1152 passing as of v1.0.0)

Two documentation surfaces, by audience. docs/ is the engineering reference — the system architecture and a per-module design doc (composer, dks, retrieval, indexer, bb, format, cli, mcp) for contributors reading the code. vault/ is the knowledge documentation — Tessellum dogfoods itself, so its concepts, how-tos, and design arguments live as typed atomic notes; start at vault/0_entry_points/entry_master_toc.md. See DEVELOPING.md for the rationale.

Compared to Adjacent Tools

License

MIT — use freely, contribute back.

Origin

Tessellum is the public release of the typed-knowledge system originally developed inside Amazon's buyer-abuse-prevention research vault. The architecture (BB ontology, Folgezettel trails, DKS protocol, CQRS thesis) was discovered through ~14 long-running Folgezettel research trails over 2024–2026.

The name Tessellum is Latin: small mosaic tile — the atomic typed unit. A vault is the mosaic.

Acknowledgments

• The Zettelkasten community (especially Sascha Fast at zettelkasten.de) for the Building Block taxonomy this work builds on
• Niklas Luhmann for proving typed atomic notes scale to ~90k connected ideas
• Tiago Forte for the PARA scheme
• The Phasespace-Labs / palinode project for independently validating the SQLite + sqlite-vec + FTS5 + RRF stack
• CQRS architects (Greg Young, Udi Dahan) for the read/write split applied here to typed knowledge