Skip to main content
Monomento indexes markdown documents into SQLite, enabling fuzzy search by topic and reference graph navigation. It stores full document content — no filesystem access needed after indexing. Published contract: monomento 0.3.2, OpenCLIs build 2026-07-18-132645, source commit 2d9f808e9cb24c7b37f065b523bc76426f27f7fb. The executing binary, registry, and public CDN metadata agree on this clean successor to the incomplete 0.3.1 publication. The installed surface has 13 operational command families, one canonical alias (path for path-search), plus docs, help, and version. Monomento does not expose a direct MCP server: higher layers such as MonoBrain, MonoFlow, and Work Wiki compose its library and database contracts. The public OpenCLIs initiate_md now contains the exact 0.3.2 packaged guide: 20,084 bytes with SHA-256 293215f99f961e4959708872063fdf9e42e591cadf68da2e99ae5fa0458f7dfd. It intentionally documents the released 13-family surface, not a source-ahead command that the installed binary cannot execute.

Bring sources into memory

Choose how to fetch and convert web pages, documents, feeds, or Wikipedia before indexing them with Monomento.

Quick Start


Memory Access Flow


Commands

Source-ahead ingest-trajectory

Current repository source at commit 73ded8644ea2749b5513a1b0061d8bcf8c44cce0 adds a 14th family:
It projects LongMemEval-V2-shaped <root>/<trajectory-id>/trajectory.json data into one Markdown document per state, using filenames such as state__t_<id>__s<NNNN>.md. The filename and frontmatter carry state_index as the span coordinate; the original step is metadata only. Each document uses the goal as H1 and Action, URL, and AXTree as H2 sections. Unchanged projected files are not rewritten, so a repeated run is idempotent at the file layer. By default the command indexes the projected corpus after writing it. --no-index performs projection only; --db, --project, --adopt, --gc, and --gc-dry-run are forwarded to the index step. The source guide has SHA-256 432b7d95a32dc03a26e8506818d706b2d492553311a55e0d15f2fcb8478a0d6b. The public 0.3.2 payload responds Unknown command: ingest-trajectory, so this section is source-head documentation, not an instruction to use the command on the current release. Every retrieval family accepts --project to select the deepest registered ancestor of the current directory. Without it, the command reads the global database. --db <path> is an index-only ownership override; it is not a general reader flag. Find documents matching a topic using trigram fuzzy matching.
Default search uses coverage × IDF with heading, body, H1, H2, and filename boosts. There are no boolean operators: AND, OR, quotes, +term, and -term are literal input. Prefer two or more distinctive words to reduce shared-trigram leakage.

path-search / path

Search canonical indexed paths rather than document content.
--contains is repeatable, and --type filters the indexed document type. This command uses a separate path index so URL/filesystem strings cannot inflate content-search scores. The installed parser accepts -n; it does not expose a --limit spelling for this command.

tree / chain

Follow reference graphs (DFS tree or BFS chain).
tree also accepts --max-children N (default 20) and --full. The latter expands repeated paths instead of deduplicating them. chain is the flat BFS view; tree is the hierarchical DFS view.

peek

Read full document structure with content previews.

read

Retrieve full content from database (no filesystem needed).
Heading selectors cover --h1 through --h6; append + to include that level and all deeper levels (--h2+, for example). With multiple documents, a trailing non-flag string is the section-name filter.

bridge

Find content-similar documents (Jaccard similarity on trigrams).

spread / resonate

spread starts from one or more exact documents and propagates activation over references; --content adds Jaccard similarity. resonate discovers lexical seeds first, then spreads and merges them. Both accept depth/decay controls; resonate --expand adds pseudo-relevance feedback.

deps / rdeps

index / stats

A successful full directory --project index registers one canonical root and its extension/exclusion policy for NIIA watcher dirty passes. --files-from applies exact newline-separated files or directory witnesses atomically under a previously stamped project or explicit DB; it never creates registration. Existing paths refresh eligible descendants, while missing paths purge their exact keys and descendants. --adopt is a fail-closed bootstrap for a legacy project corpus whose deterministic project DB already exists. It writes the extension policy with registration_catchup_pending=1, increments the generation, registers the canonical root, and returns without a directory walk. The NIIA dirty driver then drains the catch-up in batches; use the printed niia index --dirty --dirty-project <root> next step to drain immediately. It requires --project, rejects --db, --files-from, GC, and retry options, and never creates a first corpus. A fresh project must complete one full monomento index <root> --project publication. --gc-dry-run previews stale-row removal. --gc applies it, and a subdirectory walk never removes rows owned by other roots. --retry-failed clears unreadable file markers before retrying. Full and incremental writers share a per-project lock; readers fail closed while a registration catch-up generation is pending.

Storage and Freshness

Project readers choose the deepest registered ancestor of the current directory. A tombstoned, unreadable, or missing registered DB does not silently fall back to a deterministic path. Filesystem skip checks use nanosecond mtime fingerprints, while the public modified_at field remains Unix epoch seconds.

Release Platforms

The 0.3.2 build publishes six platform artifacts. latest.json and every platform row pin source commit 2d9f808e9cb24c7b37f065b523bc76426f27f7fb with source_dirty: false; this is the complete clean successor to the partial 0.3.1 publication. Archive hashes were re-read from CDN metadata and independently recomputed. Payload hashes are over the extracted executable, not the surrounding archive. Every archive also carries initiate.md, SKILL.md, initiate.json, and the deep docs/index.md topic. Their content matches the pinned 0.3.2 release source; the Windows x86-64 checkout differs only by CRLF line endings. The packaged and installed guide therefore remains the 13-family hash 293215f99f961e4959708872063fdf9e42e591cadf68da2e99ae5fa0458f7dfd, while the current repository guide is the separate source-ahead hash shown above.

Document Types

All types are parsed for headings, links, mentions, anchors, and trigram content.

Relationships