docmancer 0.6.14

Compress local documentation context for coding agents.

Your agents' memory, unified, writable, and yours.

Install | First run | What you get | Wiki

---

Your coding agents (Claude Code, Codex, Cursor, Gemini, OpenCode, Cline, Windsurf, and more) already write memory, instructions, and rules all over this machine, each locked inside its own tool. Docmancer unearths that context, recalls it at the right time, lets people and local agents write new memories, makes every atom inspectable and forgettable, and shares reviewed team knowledge through Git. The local loop is six steps:

1. Sync (docmancer memory sync): discover, redact, and extract small source-attributed atomic memories into one local SQLite index. Local, offline, no keys.
2. Recall (docmancer memory query or docmancer install <agent> --hooks): retrieve project, team, and global memory while excluding unrelated projects.
3. Remember (docmancer memory add or local MCP): write a redacted Markdown record with a stable ID and index it immediately.
4. Inspect and forget (memory list, show, and forget): review provenance, remove Docmancer-owned records, or suppress harvested atoms without editing another agent's files.
5. Capture, optionally (docmancer install <agent> --capture-hooks): extract durable atoms from local lifecycle payloads without storing raw transcripts or calling a hosted model.
6. Share after review (docmancer memory promote <id> --team): copy a memory into <repo>/.docmancer/memory/ for normal Git review. Docmancer never stages or commits it.

The same engine also does docs RAG as a secondary capability: point it at a folder of Markdown / PDF / DOCX / RTF / HTML or a docs URL (GitBook, Mintlify, generic web, GitHub) and query it the same way. A fresh install ships everything you need for the local path: SQLite FTS5 for lexical search, a static embedding model (potion-base-8M) vendored in the package so there is no large model download and no network at runtime, and sqlite-vec for dense vectors in a single local file with no daemon.

Install

pipx install docmancer    # Python 3.11, 3.12, or 3.13

If pipx picks an unsupported interpreter, pin one: pipx install docmancer --python python3.13.

First run

Two commands take you from a fresh install to recalling your agents' memory:

docmancer setup                                  # discovers and indexes the agent memory already on this machine
docmancer memory query "why did we pick Railway" # recall a past decision, offline

setup creates ~/.docmancer/ with the config and SQLite database, syncs the memory, instructions, and rules your coding agents already wrote (Claude Code, Codex, Cursor, Gemini, OpenCode, Cline, Windsurf, and more) plus repo-level CLAUDE.md / AGENTS.md / GEMINI.md, auto-detects installed agents, and installs their skill files. There is no large model download and no network at runtime: the static embedding model is vendored in the package.

Re-sync any time and see which source files were harvested and how many atoms they produced:

docmancer memory sync                # discover, redact, extract atoms, and rebuild the memory index
docmancer memory sources             # provenance: agent, type, scope, title, path, chars, atom count
docmancer memory sources --preview   # live re-harvest (what WOULD index) without writing
docmancer memory audit               # find likely secrets in source memory files

Write, inspect, forget, and share memory

Add a project decision, then inspect the stable record and atom IDs:

docmancer memory add "Production deploys run on Railway" --type decision --scope project --project "$PWD"
docmancer memory list --scope project
docmancer memory show <record-id>
docmancer memory query "where do production deploys run?" --project "$PWD"

memory forget behaves according to provenance. For a Docmancer-owned record it deletes the Markdown body and leaves a content-free tombstone. For harvested memory it leaves the other agent's source file alone and writes a local suppression record so repeated syncs do not bring the atom back.

docmancer memory forget <id> --dry-run
docmancer memory forget <id> --yes

Team memory is one editable Markdown file per atom. Add it directly when the decision is already reviewed, or promote a personal or captured memory after inspection:

docmancer memory add "Every schema change needs a rollback note" --scope team --project "$PWD"
docmancer memory promote <id> --team --project "$PWD"
git diff -- .docmancer/memory/

The packaged local MCP exposes docmancer_memory_add, docmancer_memory_list, docmancer_memory_show, docmancer_memory_forget, and docmancer_memory_promote alongside memory search. Destructive operations return a preview unless the caller explicitly sets confirm=true, and promotion accepts only a repository memory destination derived from project_path.

Run the checked-in recall regression corpus, or use the same JSONL format for a private machine-specific set:

docmancer memory eval --dataset tests/fixtures/memory-eval-synthetic.jsonl
docmancer memory eval --dataset memory-eval-private.jsonl --format json

Want docs RAG too? The same engine indexes documentation:

docmancer ingest ./docs                             # index local files
docmancer add https://docs.pytest.org               # or a docs URL
docmancer query "How do I parametrize a fixture?"   # hybrid search across the docs index

Automatic recall with hooks

Syncing gives you one searchable atomic-memory index. Hooks make that index useful while you work: on SessionStart and UserPromptSubmit, docmancer runs a fast local query and injects only the top relevant source-backed atoms into Claude Code or Codex context. Most prompts inject nothing. Weak matches stay silent, and a per-session cache avoids repeating the same memory every turn.

docmancer install claude-code --hooks
docmancer install codex --hooks

Hooks are local and bounded. They read hook JSON from stdin, query the existing local atomic-memory index, and output hook-compatible additionalContext only when relevant matches clear the threshold. They never call OpenRouter or another provider. If the hook is slow, malformed, or has nothing useful to add, it exits successfully with no output so the agent turn continues normally. The internal hook budget defaults to 1,000 ms and can be adjusted with DOCMANCER_HOOK_TIMEOUT_MS. Remove hooks with:

docmancer remove claude-code --hooks
docmancer remove codex --hooks

Codex hooks use the documented ~/.codex/hooks.json hook shape and emit hookSpecificOutput.additionalContext on stdout. Codex may require you to review and trust the installed hook through /hooks.

Capture is a separate opt-in. Claude Code capture uses PostCompact and SessionEnd; Codex uses PreCompact and Stop. Capture redacts first, reads only a bounded transcript tail when needed, stores extracted durable atoms rather than raw transcripts, skips malformed or low-value events, and never blocks an agent turn.

docmancer install claude-code --capture-hooks
docmancer install codex --capture-hooks
docmancer remove claude-code --capture-hooks
docmancer remove codex --capture-hooks

Optional consolidation and apply

Consolidation is optional maintenance, not the main memory-transfer path. docmancer memory consolidate sends selected, privacy-redacted atomic memories to OpenRouter and returns a review-only draft: deduplicated, grouped into compact sections, with conflicts surfaced as warnings instead of silently resolved.

export OPENROUTER_API_KEY=...
docmancer memory consolidate \
  --provider openrouter \
  --model openai/gpt-4.1-nano \
  --query "deployment and infra decisions" \
  --output master-memory-draft.md \
  --yes

Materialize selected atomic memories into an agent's always-loaded file only if you want a compact stable summary:

docmancer memory apply --agent codex   # renders atomic memories from the local index
docmancer memory apply --agent codex --dry-run   # preview the diff first

apply is local and keyless. It writes only inside a clearly delimited managed block, takes a timestamped backup first, and never touches your own surrounding content. --remove strips the block for a clean uninstall. Passing --from draft.md still applies a reviewed consolidation draft for older workflows, but the default path renders atomic memories from the index. (docmancer install codex / claude-code also inject a short recall instruction into the same files, in their own managed block.)

Use --timeout or DOCMANCER_OPENROUTER_TIMEOUT_SECONDS to bound each OpenRouter request, with a finite 180 second default and 0 for the provider default. Provider-backed commands fail gracefully with concise messages, print a provider-use notice before the first call, run a preflight before large memory payloads, log each request before sending it, and run secret redaction before any text leaves your machine. See the Configuration and Commands pages for details.

What you get

A complete local memory loop. docmancer memory sync unifies existing agent memory; add, list, show, and forget make the layer writable and inspectable; project-aware recall isolates workspaces; optional capture extracts durable local atoms; team promotion produces reviewable Git files. The local path uploads nothing.

Automatic recall in Claude Code and Codex. docmancer install claude-code --hooks and docmancer install codex --hooks inject relevant atomic memories at session start and before matching prompts. The hook path is local, bounded, source-attributed, and silent when nothing relevant clears the threshold.

Memory audit. docmancer memory audit scans harvested source memory before redaction and reports likely secrets with masked excerpts, short source labels, line numbers, and a clear next action. It is local and read-only. It reports and locates only; it never edits another tool's files.

Optional consolidation. docmancer memory consolidate --provider openrouter turns selected atomic memories into one review-only master-memory draft when OPENROUTER_API_KEY is configured. docmancer memory apply can render atomic memories into an agent's always-loaded file, with diff preview, backups, and undo.

Readable and writable over MCP. The packaged docmancer-mcp stdio server exposes local memory search, add, list, show, forget, and promote tools plus docs search. Forget and promote require explicit confirmation. docmancer mcp install codex (or claude-code, claude-desktop) wires it up; optional OpenRouter consolidation appears only when OPENROUTER_API_KEY is set. Requires the mcp extra.

Hybrid search by default. query and memory query fan out across SQLite FTS5 (lexical, BM25-reranked) and dense vectors from a vendored static model (potion-base-8M) in sqlite-vec, then fuse results with Reciprocal Rank Fusion. Sparse (SPLADE) signals are available on the optional heavy Qdrant backend. The token budget keeps responses small so your agent has room for actual work:

Context pack: ~900 tokens vs ~4800 raw docs tokens (81.2% less docs overhead, 5.33x agentic runway)

No large model download, offline at runtime. The static embedding model ships inside the wheel, so there are no API keys and no network needed to embed or query. Optional OpenAI / Voyage / Cohere providers exist for users who explicitly configure cloud embeddings.

Where your data lives and how to remove it

The local surfaces have separate, inspectable locations:

• ~/.docmancer/memory.db and its co-located sqlite-vec file are the rebuildable search index.
• ~/.docmancer/memories/*.md contains personal, project, manual, and captured records as Markdown with YAML frontmatter.
• <repo>/.docmancer/memory/*.md contains free, MIT-licensed team memory intended for Git review.
• ~/.docmancer/memory-tombstones.json contains content-free suppression identifiers and hashes.
• The atom extraction cache and hook dedupe cache live under ~/.docmancer/; neither is a transcript archive.
• OpenRouter receives privacy-redacted selected memory only when you explicitly run optional consolidation. Local sync, CRUD, MCP, hooks, capture, and evaluation do not call it.

docmancer memory clear deletes the rebuildable index but deliberately leaves durable Markdown records and tombstones. docmancer remove <agent> --hooks removes all Docmancer hooks. Use docmancer remove <agent> --capture-hooks when you want to remove capture while leaving recall installed. There is no telemetry and no phone-home.

Inspectable. Every section is written to ~/.docmancer/extracted/ as Markdown plus JSON. docmancer inspect shows index stats. docmancer query --explain shows which signal (lexical / dense / sparse) placed each result.

Agent integration built in. docmancer setup drops skill files for Claude Code, Cursor, Codex, Cline, Claude Desktop, Gemini, GitHub Copilot, and OpenCode. For Claude Code and Codex it also injects a memory-recall instruction into the always-loaded CLAUDE.md / ~/.codex/AGENTS.md (a managed block), so manual pull recall still works when hooks are not installed.

Where to next

The wiki is the authoritative reference for everything else. Pick a page based on what you need:

Page	When to read it
Commands	The memory loop, docs commands, and advanced maintenance
Configuration	All YAML keys, env vars, and the API-key reference
Architecture	How the memory harness, ingest, and hybrid retrieval work
Supported Sources	What file formats and URL providers are covered
Install Targets	Where each agent's skill file lands
Troubleshooting	Common errors and fixes

Wiki home | Changelog | PyPI