Your agents' memory. Unified and yours.
Docmancer syncs Claude Code, Codex, Cursor, Gemini, and more into one local memory index. Recall past decisions offline, then optionally use Mistral AI to consolidate everything into a reviewed memory you can carry across agents.
Works with
Claude Code
Cursor
Codex
Cline
OpenCode
GeminiClaude DesktopPowered by Mistral AI
From scattered agent memory to a reviewed master file.
Syncing gives you one searchable index, and consolidation turns that pile into a single coherent memory. Run docmancer memory consolidate with no arguments and Mistral AI returns a deduplicated, review-only draft with conflicts flagged rather than silently resolved. Once you have read it, docmancer memory apply bakes that draft into an agent file so the same memory loads across all of them.
Structured extraction
Mistral structured outputs pull durable facts (decisions, conventions, constraints) out of raw agent memory, with evidence and confidence on each one.
Review-only consolidation
A Mistral chat model merges duplicates and groups everything into sections. Conflicts surface as warnings instead of being silently resolved. It never edits your agent files.
Private and explicit
Secrets are redacted before any text leaves your machine, you confirm the cloud call once, and with no key set the command exits cleanly without touching your local index.
Features
Built for serious retrieval. Local by default.
01 · Finds your harnesses
The memory your agents wrote, unified.
One sync discovers memory, instructions, and rules across your coding-agent harnesses, then indexes them into a local store you own. docmancer memory sources shows exactly what came from where.
02 · Stays local
Local first. Cloud only when you ask.
Sync and recall run locally with vendored static embeddings, SQLite FTS5, and sqlite-vec. Mistral is optional, key-gated, and only receives selected memory after redaction and confirmation.
03 · Search that actually works
Lexical and dense, fused into one.
Docmancer combines exact-term search with local dense vectors, then fuses the rankings with Reciprocal Rank Fusion. Exact names and paraphrases both have a path to the top.
04 · Answers with receipts
Every result carries its source.
Results include source paths or URLs, so your agent can cite what it used and you can inspect the extracted Markdown whenever you need to.
05 · Drops into your agent
Your editor already knows how to call it.
docmancer setup detects installed agents and writes their skill files. Prefer MCP? The packaged docmancer-mcp server installs into Codex, Claude Code, or Claude Desktop.
Process
How it works.
Three steps take you from scattered agent memory to a local recall loop. The default path is keyless and offline; Mistral only enters when you ask it to consolidate a reviewed draft.
Set up
One command creates the local memory index, syncs the memory and instructions your coding-agent harnesses already wrote, and wires docmancer into the agents installed on your machine.
$ docmancer setupRecall
Ask for the decision, rule, setup note, or past context you need. Docmancer searches the local hybrid index and returns compact results with source paths attached.
$ docmancer memory query "why Railway"Consolidate
When the pile gets noisy, run consolidate with no arguments to turn redacted memory into a review-only draft with optional Mistral AI, then apply the reviewed result back into an agent file.
$ docmancer memory consolidateFAQ
Questions, answered.
The default sync and recall path runs on your machine. The memory index is stored in local SQLite-backed files under ~/.docmancer, and embeddings come from a static model vendored inside the package. Optional Mistral commands send selected, privacy-redacted memory text only after confirmation.
Stop losing context between agents.
One pipx line. About three minutes. Sync the memory your agents already wrote, recall it locally, and optionally consolidate it with Mistral AI.